Insects visually solve an array of complex problems including the detection and tracking of fast moving prey (Wiederman et al., 2017), long-distance navigation (Wehner, 2020), and even three-dimensional (3D) depth estimation (Nityananda et al., 2018). These capabilities are realised using a sensory apparatus that is fundamentally different from those of mammals. Therefore revealing the functional properties of the insect visual system offers insights for biologists as well as inspiration for engineers looking to develop novel artificial imaging systems (Land and Fernald, 1992; Land, 1997; Arendt, 2003; Song et al., 2013).

Arthropods possess two primary visual sensors known as compound eyes. Each eye is constructed from a patchwork of self-contained light-sensing structures known as ommatidia, each featuring a lens, a light guide, and a cluster of photosensitive cells (Figure 1a). Ommatidia are physically interlocked with their neighbours, together forming a bulbous outer structure (the compound eye itself) that can vary in size, shape, and curvature, offering a range of adaptations for particular tasks and environments (Land and Nilsson, 2002; Figure 1b). The properties of the lens and photo-receptors are fixed for individual ommatidia but can vary across regions of an individual’s eye (Meyer and Labhart, 1993) as well as between individuals of different castes (Collett and Land, 1975) and species (Land, 1989). This arrangement of independent, interlocked light sensing elements and lenses differs greatly from the mammalian system that utilises a single lens to project a high-resolution image onto a retina.

Figure 1

Download asset Open asset

Creation of compound eye models (CEMs), in both hardware or software, is a well-established mechanism to explore the information provided by compound eyes and assess their impact on behaviour. Insights derived from this methodology include demonstration of the benefits of a wide field of view (FOV) and low resolution for visual navigation (Zeil et al., 2003; Vardy and Moller, 2005; Mangan and Webb, 2009; Wystrach et al., 2016), and the role played by non-visible light sensing for place recognition (Möller, 2002; Stone et al., 2006; Differt and Möller, 2016) and direction sensing (Lambrinos et al., 1997; Gkanias et al., 2019). Yet, simulated CEMs tend to suffer from a common design shortcoming that limits their ability to accurately replicate insect vision. Specifically, despite differences in the sampling techniques used (e.g. see Mangan and Webb, 2009; Baddeley et al., 2012 for custom sampling approaches, Neumann, 2002; Basten and Mallot, 2010 for rendering based cubemapping CEMs, and Polster et al., 2018 for ray-casting methods), all contemporary CEMs sample from a single viewpoint. In contrast, the distributed arrangement of ommatidia on distinct 3D eye surfaces provides insects with a multi-viewpoint system that generates different information for different eye shapes.

To facilitate exploration of such features, e.g., the placement of ommatidia on arbitrary surfaces, an ideal rendering system would allow light to be sampled from different 3D locations through individually configured ommatidia replicating the complex structure of real compound eyes. High-fidelity compound-vision rendering engines with some of these features (though notably slower than real time) were developed previously (Giger, 1996; Collins, 1998; Polster et al., 2018) but were not widely adopted. Difficulties arising as a result of the computational complexity (and so execution time) of CEMs are diminishing as dedicated ray-casting hardware emerges that allows for the capture of visual data from multiple locations in parallel at high speed (e.g. Nvidia [Santa Clara, California, United States] RTX series GPUs [Purcell et al., 2005; Burgess, 2020]). Such systems present an ideal tool tferreo replicate insect vision in unprecedented accuracy at the speeds needed for effective exploration of the compound eye design space itself. As outlined in Millward et al., 2020, a next-generation insect eye renderer should:

1. Allow for the arrangement of an arbitrary number of ommatidia at arbitrary 3D points.

2. Allow for the configuration of individual ommatidial properties (e.g. lens acceptance angle).

3. Perform beyond real time to allow exploration of the design space.

This paper presents a ray-casting-based renderer, CompoundRay, that leverages modern hardware-accelerated ray-tracing (RT) graphics pipelines and fulfils all three of these criteria, allowing researchers in the fields of compound vision and biorobotics to quickly explore the impact varying eye designs have on an autonomous agent’s ability to perceive the world.

The analysis that follows first assesses the performance of CompoundRay with respect to the three criteria defined in the ‘Introduction’, before showing its utility in an example experiment. Performance is benchmarked in two 3D environments: a lab environment inspired by those used in insect cognitive experiments (Ofstad et al., 2011); and a 3D scan of a real-world natural environment covering an area of a number of square kilometers. The natural environment was constructed using structure-from-motion photogrammetry using photos captured from a drone over rural Harpenden, UK (3D model subject to upcoming publication, available via contact of Dr. Joe Woodgate, Queen Mary University of London).

Criterion 1: arbitrary arrangements of ommatidia

Criterion 1 states that the renderer must support arbitrary arrangements of ommatidia within 3D space. Figure 6 shows two different environments rendered using three differing eye designs, from a simple spherical model with spherical uniformly distributed ommatidia that more closely aligns with current compound eye rendering methods, to a hypothetical arbitrary arrangement similar to that of the now long-extinct Erbenochile erbeni. While Figure 6a&b is still possible to simulate using traditional projection-based methods due to their ommatidial axes converging onto one central spot, 6 c demonstrates an eye with multiple unique focal points across an irregular surface which must be rendered from the viewpoint of each ommatidium independently.

Figure 6

Download asset Open asset

Criterion 2: inhomogeneous ommatidial properties

Criterion 2 states that it should be possible to specify heterogeneous ommatidial optical properties across an eye, such as the enlarged facets as found in robberflies (Wardill et al., 2017). As demonstrated in Figure 7, the renderer achieves this by allowing for shaping of the acceptance cone of each ommatidium. In doing so, Figure 7 also demonstrates the importance of heterogeneous ommatidial configurations.

Figure 7

Download asset Open asset

Figure 7a&b shows a non-uniform eye design with homogeneous, globally identical, ommatidial acceptance angles. It can be seen that in Figure 7b, where the acceptance angle is lower, aliasing occurs as objects are only partially observed through any given ommatidium, causing blind spots to form, introducing noise into the image. Conversely, in Figure 7a, where the acceptance angle is large, the image becomes blurry, with each ommatidium oversampling portions of its neighbour’s cone of vision. Figure 7c, however, demonstrates a heterogeneous ommatidial distribution, in which the acceptance angles toward the dorsal and ventral regions of the eye are larger compared with the angles of those around the eye’s horizontal acute zone. This mirrors what is seen in nature, where ommatidial FOV is seen to vary in order to encompass the entirety of an insect’s spherical visual field (Land and Nilsson, 2002). As can be seen in the generated image, this is important as it appropriately avoids creating blind-spots in the dorsal and ventral regions while also avoiding oversampling (blurring) in the horizontally acute region, resulting in a much clearer picture of the observed environment (before the visual data is passed into any insect visual processing neural circuits) while also minimising the required number of ommatidia. The data presented in Figures 6 and 7 clearly demonstrates differences in visual information provided by different ommatidial layouts, reinforcing the need for realistic eye modelling methods.

Criterion 3: speed

Minimum sample count

Criterion 3 states that to rapidly explore the information content of various eye designs any such rendering system should be able to perform at real time or faster. However, speed of the system is dependent on the number of rays required to render a given scene from a given eye; higher total ray counts will require more compute power. As discussed in the section ‘Modelling individual ommatidia’, per-ommatidial sampling choice can have a dramatic impact on the image generated, meaning that lower ray counts – while increasing speed – will decrease the accuracy of the image produced. As a result, it is important to be able to define some minimum number of samples required for any given eye in any given environment in a reliable and repeatable manner. Previous works Polster et al., 2018 have attempted to establish a baseline number of samples required per ommatidium before additional samples become redundant only via qualitative visual analysis of the results produced. Here we perform quantitative analysis of images produced by the renderer in order to more definitively define this baseline (Figure 8).

Figure 8

Download asset Open asset

Unlike previous works, the temporally stochastic nature of CompoundRays’ sampling approach can be used to aid the measurement of impact that sampling rate has on the final image. By measuring the per-ommatidium spread (here, standard deviation) of the Euclidean distance of the received colour as plotted in RGB colour space over a range of frames captured from a static compound camera, the precision of the light sampling function approximation can be measured. To control for the varying FOV of each ommatidium (which impacts the number of samples required, as larger FOVs require more sampling rays to accurately capture the ommatidium’s sampling domain), we normalize the measure of spread recorded at each ommatidium against its degree of coverage, in steradians, expressing the solid angle subtended by the ommatidium’s FOV at the eye’s centre. That is, a steradian measures the solid angle within a sphere, in the same same way that a radian measures 2D angles within a circle.

Figure 8 shows this measure of variance within the compound eye across a range of sample-rays-per-ommatidium, from 1 to 700. As the number of samples increases, the variance decreases logarithmically. We propose iteratively increasing samples until the maximum variance falls below a defined threshold value. Here, we use a threshold of 1%, meaning that the maximum expected standard deviation of the most deviant ommatidium on an eye (normalised to one steradian) should be no greater than 1% of the maximum difference between two received colours, here defined as the length of vector [255, 255, 255]. This method allows for standardised, eye-independent configuration of per-ommatidial sampling rate.

However, the required number of samples per eye is highly positively correlated with the visual frequency and intensity of the simulated environment at any given point, with locations exhibiting high contrast seen from a distance requiring higher sampling rates in order to properly resolve. This can be seen in the skyline in Figure 8, which remains dominant throughout the majority of the sampling rates, consistently forming the source of the highest spread. Figure 9 shows how the perceived variance changes over two example environments. The varying environment presents an obstacle in our attempts to standardise sampling rates, as deriving a suitable minimum sampling rate at one point in an environment – e.g., where the visual frequency was, on average, low – could lead to aliasing occurring in more visually complex areas of the same environment, potentially introducing bias for or against certain areas if analysis on the generated data were sensitive to the variance introduced.

Figure 9

Download asset Open asset

Ultimately, the varying visual complexity poses a challenge. In an ideal scenario maximum per-ommatidium frame-to-frame variation caused by visual complexity should be a controllable factor. In the examples presented in this paper, we worked to minimise this variation to a point at which the most variable ommatidial response had a standard deviation of 1% or less of the total maximal difference in colour. One way of achieving this would be to keep a running measure of the standard deviation of each ommatidium and dynamically adapt sampling rates to ensure that no ommatidium had such a high sampling spread (adaptive antialiasing). In this work, however, we define a minimum sampling count for a given environment by performing a search for that environment’s point of highest visual complexity (indicated by highest per-ommatidium deviation) – with this location found, sampling rate can be increased until deviation decreases to acceptable limits (in this case, within 1%). Figure 9 shows plots of the relative variation in standard deviation over the lab and (partial) natural environments.

Performance

Figure 10 shows the rendering speed, in frames per second, against the number of sample rays being emitted in each frame when running using a selection of high-end consumer-grade graphics cards and our two sample environments. Marked on the graph are the total per-frame sample counts required by common insect eyes, chosen to encourage a maximum per-ommatidium standard deviation of 1%.

Figure 10

Download asset Open asset

As can be seen, performance improves significantly when the renderer is able to utilise the next-generation RT hardware available in the newer RTX series of NVidia graphics cards, consistently outperforming older cards in the larger natural environment and outperforming in the smaller lab environment when using more than 30,000 rays. Counterintuitively, the previous generation of NVidia graphics cards was found to outperform the RTX series in the (relatively small) lab environment at lower ray counts (below 30,000). This appears to be due to static overhead involved with the use of the RTX series’ RT cores, which may also be the cause of the small positive trend seen between 1000 and 8000 samples when using the 2080Ti. Another potentially counterintuitive result that can be seen here is the higher performance of the dragonfly eye in the natural (high polygon) environment versus in the low-polygon laboratory environment. This is because the laboratory environment produces significantly higher visual contrast (emphasised in Figure 9) due to its stark black/white colour scheme, resulting in a requirement of almost double the number of rays per steradian to target 1% standard deviation, resulting in higher ray counts overall, particularly in models with a high number of ommatidia.

Example experiment: Apis mellifera visual field comparison

Leveraging the speed of CompoundRay, an example data-driven comparative analysis was performed between a realistic eye model and two simplified derivatives. We note that the data outcomes should be considered as pilot data, with the focus being on proving the tool and its usability. As laid out in the section ‘Example inter-eye comparison method’, this experiment used an MLP network to estimate the relative position of a 2 mm sphere within a 50 mm³ test area using only the visual information provided from each of the eye models, with neural network performance acting as a measure of eye configuration utility. The models were split into the ‘real’ eye model (Figure 5i), an eye pair based directly on the measurements of Baird and Taylor, 2017; the ‘split’ eye model (Figure 5ii), an eye pair retaining per-ommatidial headings (ommatidial axial directions), but sampled from the centre points of each eye (as per approaches common in insect binocular vision study); and finally the ‘single’ eye model, an eye pair sampled from a single point between the two eyes (Figure 5iii). We hypothesise that the ‘real’ eye design will outperform the ‘split’ and ‘single’ eye models by increasing margins, as for each of the latter two designs the data relating to the surface shape and relative positions of the eyes is increasingly reduced.

In terms of total error over the validation set, we observed that – as expected – the ‘real’ eye design appears to offer a higher utility (lower total error when trying to locate the relative position of the sphere), ultimately decreasing its network’s validation error quicker and to a lower minimum after 100 epochs when compared to both the ‘split’ and ‘single’ eye designs (Figure 11a). Interestingly, despite an initial slow in learning rate, the ‘single’ eye design eventually achieved a similar utility score on the validation set than that of the ‘split’ eye design.

Figure 11

Download asset Open asset

To qualitatively assess the performance of each eye across the entire 50 mm³ sampling cube, a further 1,000,000 samples were taken at regular intervals forming an error volume for each eye design by measuring the error between the trained neural network’s relative position and the real relative position (the error volume for the ‘real’ eye is pictured in Figure 5c). In this case, to demonstrate versatility, CompoundRay was used in real time to generate these views, rather than simply being used in an offline manner to generate these benchmarking datasets.

The most obvious region of high error is that outside of the insect’s FOV. As these eyes were designed with uniform average acceptance angles, this is larger than might be expected of a real bee eye. Comparing the ‘real’ eye to its ‘single’ and ‘split’ counterparts by means of comparing the absolute difference between their error volumes (creating the difference volumes shown from top-down in Figure 11b&c), patterns emerge. Seemingly, the most significant region of difference between the ‘real’ eye model and the ‘single’ eye model can be seen around the edges of the eye’s field of vision. As this difference follows the entirety of the edge of the eye’s field of vision, this appears to be likely due to parallax – the ‘real’ eye samples from two points, allowing it slightly extended range of vision compared to that of the ‘single’ eye model, resulting in slightly different ranges of FOV between the two, with this difference increasing the further from the eye the point is. In the ‘real’ vs ‘split’ comparison image, however, there are two regions of higher error (highlighted in red circles in Figure 11c) toward the forward-facing regions of the eye’s vision. These indicate a region of improved visual information when using a full eye surface as compared to simple two-point sampling of the environment.

This paper has introduced a new compound eye perspective rendering software, CompoundRay, that addresses the need for a fast, compound eye perspective rendering pipeline (Stürzl et al., 2015; Taylor and Baird, 2017; Millward et al., 2020). In particular, the tool supports arbitrary instantiation of heterogeneous ommatidia at any place in space and can perform rendering tasks at speeds in the order of thousands of frames per second using consumer-grade graphics hardware. We have demonstrated the utility afforded by the software and highlighted the importance of using higher-fidelity compound eye perspective rendering systems by demonstrating the visual differences introduced by altering these traits. It has also set a grounding for reasonable use of the software to ensure reproducibility and reduce biasing introduced as a result of the variety present in simulated environments and compound eye designs while warning of potential dangers that might arise during data analysis. With the introduction of the CompoundRay rendering library, it is possible to simulate in a timely manner the visual experience of structurally diverse compound eyes in complex environments, digitally captured or otherwise manually constructed. CompoundRay is provided as an open-sourced library (https://github.com/BrainsOnBoard/compound-ray).

This work lays the technical foundations for future research into elements of compound eye design that have previously been given little consideration. In particular, CompoundRay can act as the previously missing step to integrate recently catalogued real eye designs (Baird and Taylor, 2017; Bagheri et al., 2020) into mapped insect environments (Stürzl et al., 2015; Risse et al., 2018) to explore the information provided to insects in their natural surroundings. Similarly, the ability to configure ommatidial properties individually and with respect to eye surface shape enables new investigations into the benefits of asymmetric compound eyes such as those found in robber flies (Wardill et al., 2017) and crabs (Zeil et al., 1986). Furthermore, we see opportunities to model the recently reported microsaccadic sampling in fruit flies (Juusola et al., 2017), as well as aid with similar vision-centric neurological studies (Viollet, 2014; Wystrach et al., 2016; Kemppainen et al., 2021).

Our example study demonstrated the utility of CompoundRay for this strand of research. Beyond its ability to replicate insect vision in unprecedented detail, its image generation speeds facilitate use of contemporary data-driven methods to explore the eye design space. For example, we used over 3.3 million training images, which CompoundRay rendered in only 2 hr. The insights gained from more thorough analysis of the insect visual perspective—and its design relative to visual feature extraction—will help guide the development of artificial visual systems by considering not only visual post-processing steps but also the intrinsic structure and design of the image retrieval system itself. CompoundRay is explicitly designed to be highly configurable to allow members of the research community to pursue their specific research questions.

In our example study, we compared eye designs using an MLP network as a utility function, bypassing the need to directly compare the images generated by two differing eye designs. This was done as the task of directly comparing two dissimilar eye designs – either analytically using a comparison program or qualitatively via human observation – is non-trivial. Compound eyes are inherently different from our own, and the CompoundRay rendering system allows for large-scale data collection on arbitrary compound surfaces. We note that with this comes the risk of anthropomorphising the data received when directly observing it, or placing bias on one area over another when comparing programmatically. At its base level, the eye data gathered is most ‘pure’ when considered as a single vector of ommatidial samples, much in the same way that the neural systems of any insect will interpret the data. In contrast, all the images that we have presented here have been projected onto a 2D plane in order to be displayed. Great care must be taken when considering these projections, as they can introduce biases that may be easy to overlook. For instance, in Figure 6a&b, orientation-wise spherical projection mapping was used to artificially place the ommatidia across the 2D spaces, forming panoramic Voronoi plots of observed light at each ommatidium. In these projections, much as is seen in 2D projections of Earth, the points at the poles of the image are artificially inflated, giving a bias in terms of raw surface area affected toward ommatidia present in the ventral and dorsal regions of the eye. Not only do these visual distortions warp the image produced, but if that image is used as the basis for algorithmic comparison (e.g., as the input to an image difference function [Philippides et al., 2011]) then that too, will be effected by these projection biases. In the case where direct comparison between two eyes with entirely dissimilar surfaces is attempted, the problems can be magnified further, as the projections of each eye will introduce biases that may interact in unexpected ways.

These projection biases highlight how fundamentally difficult it is to directly compare eye designs that do not share an underlying surface structure, unless those surfaces happen to be affine transformations of one another (they are affinely equivalent). For any surfaces that are affinely equivalent, the surface itself can be used as a projection surface. In this way, for instance, the outputs of two cubic eyes could be compared to each other by performing all comparison calculations on the cube’s surface, interpolating between points where an ommatidia exists on one eye but not another. It is trivial to directly compare compound eye images from eyes with the same structure in terms of ommatidial placement, and eyes with surfaces that are affinely equivalent can also be compared; however, further precautions must be taken when comparing the images from two or more eyes that do not have compatible projection surfaces so as not to unduly bias a portion of one surface over any other. If direct comparison between two eye designs is required, and their surfaces are not affine transformations of each other, then intermediate surfaces to perform calculations over must be found within the interpolation from one surface to the other, with a significant amount of care taken to reduce as much as possible projection artefacts forming between the two. However, we recommend instead taking a proxy metric as demonstrated here (using a localisation task) to measure differences between eye-accessible information content.

By basing CompoundRay around the conceptually simple approach of ray-casting and adhering to open standards, the rendering system has been designed with open development in mind. In the future, the library could be extended to further enhance biological realism by adding non-visible spectral (e.g. ultraviolet) lighting (Möller, 2002; Stone et al., 2006; Differt and Möller, 2015) and polarisation (e.g.Gkanias et al., 2019) sensitivities, as well as more realistic photo-receptor response characteristics (e.g. Song et al., 2009). Similarly, new technologies could be rapidly evaluated by simulating the input to novel imaging systems (e.g. dynamic; Viollet and Franceschini, 2010 or heterogeneous pixel arrays) and processing pipelines (e.g. event-based retinas; Gallego et al., 2022). In addition, environmental features such as shadow casting, reflection, and refractions present obvious extensions as well the introduction of features such as adaptive anti-aliasing and increasing the projections available to users. Currently, CompoundRay returns all output images in an eight-bit colourspace, which aligns with most contemporary display equipment. However, we note that the underlying GPU implementation operates on 32-bit floating-point variables which if exposed to the high-level Python API could improve the system’s ability to represent visual sensors that experience a wider bandwidth of light (hyperspectral visual systems). It is hoped that as further efforts are conducted to map the eye structures of insects, CompoundRay will serve as a key tool in uncovering the intricacies of these eye designs in a modern, data-driven way.

The manuscript is a computational study, with all modelling code and data accessible on GitHub at https://github.com/ManganLab/eye-renderer. Use of the natural environment was kindly provided by Dr. JoeWoodgate, Queen Mary University of London and is subject to upcoming publication. As such, instead included in the CompoundRay repository is a stand-in natural 3D terrain model. As all models are used for demonstrative purpose, this stand-in model offers little difference to the natural model used, bar it's subjectively lower-quality aesthetics.

1. Conference

   1. Appel A

   (1968) Some techniques for shading machine renderings of solids the April 30--May 2, 1968

   spring joint computer conference.

   https://doi.org/10.1145/1468075.1468082

   • Google Scholar
2. 1. Arendt D

   (2003)

   Evolution of eyes and photoreceptor cell types

   The International Journal of Developmental Biology 47:563–571.

   • PubMed
   • Google Scholar
3. 1. Baddeley B
   2. Graham P
   3. Husbands P
   4. Philippides A

   (2012) A model of ant route navigation driven by scene familiarity

   PLOS Computational Biology 8:e1002336.

   https://doi.org/10.1371/journal.pcbi.1002336

   • PubMed
   • Google Scholar
4. 1. Bagheri ZM
   2. Jessop AL
   3. Kato S
   4. Partridge JC
   5. Shaw J
   6. Ogawa Y
   7. Hemmi JM

   (2020) A new method for mapping spatial resolution in compound eyes suggests two visual streaks in fiddler crabs

   The Journal of Experimental Biology 223:jeb210195.

   https://doi.org/10.1242/jeb.210195

   • PubMed
   • Google Scholar
5. 1. Baird E
   2. Taylor G

   (2017) X-Ray micro computed-tomography

   Current Biology 27:R289–R291.

   https://doi.org/10.1016/j.cub.2017.01.066

   • PubMed
   • Google Scholar
6. 1. Basten K
   2. Mallot HA

   (2010) Simulated visual homing in desert ant natural environments: efficiency of skyline cues

   Biol Cybern 102:413–425.

   https://doi.org/10.1007/s00422-010-0375-9

   • PubMed
   • Google Scholar
7. 1. Burgess J

   (2020) Rtx on—the NVIDIA Turing GpU

   IEEE Micro 40:36–44.

   https://doi.org/10.1109/MM.2020.2971677

   • Google Scholar
8. 1. Christensen P
   2. Fong J
   3. Shade J
   4. Wooten W
   5. Schubert B
   6. Kensler A
   7. Friedman S
   8. Kilpatrick C
   9. Ramshaw C
   10. Bannister M
   11. Rayner B
   12. Brouillat J
   13. Liani M

   (2018) RenderMan: an advanced path-tracing architecture for movie rendering

   ACM Transactions on Graphics 37:1–21.

   https://doi.org/10.1145/3182162

   • Google Scholar
9. 1. Collett TS
   2. Land MF

   (1975) Visual control of flight behaviour in the hoverflysyritta pipiens L

   Journal of Comparative Physiology ? A 99:1–66.

   https://doi.org/10.1007/BF01464710

   • Google Scholar
10. Book

    1. Collins S

    (1998) Reconstructing the visual field of compound eyes

    In: Rocamora SP, editors. Eurographics Rendering Workshop. Springer Wien. pp. 81–92.

    https://doi.org/10.1007/978-3-7091-6858-5

    • Google Scholar
11. 1. Differt D
    2. Möller R

    (2015) Insect models of illumination-invariant skyline extraction from UV and green channels

    Journal of Theoretical Biology 380:444–462.

    https://doi.org/10.1016/j.jtbi.2015.06.020

    • PubMed
    • Google Scholar
12. 1. Differt D
    2. Möller R

    (2016) Spectral skyline separation: extended landmark databases and panoramic imaging

    Sensors 16:10.

    https://doi.org/10.3390/s16101614

    • PubMed
    • Google Scholar
13. 1. Franz MO
    2. Schölkopf B
    3. Mallot HA
    4. Bülthoff HH

    (1998) Where did I take that snapshot? scene-based homing by image matching

    Biol Cybern 79:191–202.

    https://doi.org/10.1007/s004220050470

    • Google Scholar
14. 1. Gallego G
    2. Delbruck T
    3. Orchard G
    4. Bartolozzi C
    5. Taba B
    6. Censi A
    7. Leutenegger S
    8. Davison AJ
    9. Conradt J
    10. Daniilidis K
    11. Scaramuzza D

    (2022) Event-based vision: A survey

    IEEE Transactions on Pattern Analysis and Machine Intelligence 44:154–180.

    https://doi.org/10.1109/TPAMI.2020.3008413

    • PubMed
    • Google Scholar
15. Thesis

    1. Giger AD

    (1996)

    PhD thesis: Honeybee vision: analysis of pattern orientation

    The Australian National University.

    • Google Scholar
16. 1. Gkanias E
    2. Risse B
    3. Mangan M
    4. Webb B

    (2019) From skylight input to behavioural output: a computational model of the insect polarised light COMPASS

    PLOS Computational Biology 15:e1007123.

    https://doi.org/10.1371/journal.pcbi.1007123

    • PubMed
    • Google Scholar
17. Report

    1. Golmant N
    2. Vemuri N
    3. Yao Z
    4. Feinberg V
    5. Gholami A
    6. Rothauge K
    7. Mahoney MW
    8. Gonzalez J

    (2018)

    On The Computational Inefficiency Of Large Batch Sizes For Stochastic Gradient Descent Technical report

    ICLR.

    • Google Scholar
18. 1. Greiner B
    2. Ribi WA
    3. Warrant EJ

    (2004) Retinal and optical adaptations for nocturnal vision in the halictid bee megalopta genalis

    Cell and Tissue Research 316:377–390.

    https://doi.org/10.1007/s00441-004-0883-9

    • PubMed
    • Google Scholar
19. 1. Harris CR
    2. Millman KJ
    3. van der Walt SJ
    4. Gommers R
    5. Virtanen P
    6. Cournapeau D
    7. Wieser E
    8. Taylor J
    9. Berg S
    10. Smith NJ
    11. Kern R
    12. Picus M
    13. Hoyer S
    14. van Kerkwijk MH
    15. Brett M
    16. Haldane A
    17. Del Río JF
    18. Wiebe M
    19. Peterson P
    20. Gérard-Marchant P
    21. Sheppard K
    22. Reddy T
    23. Weckesser W
    24. Abbasi H
    25. Gohlke C
    26. Oliphant TE

    (2020) Array programming with numpy

    Nature 585:357–362.

    https://doi.org/10.1038/s41586-020-2649-2

    • PubMed
    • Google Scholar
20. 1. Juusola M
    2. French AS

    (1997) Visual acuity for moving objects in first- and second-order neurons of the fly compound eye

    Journal of Neurophysiology 77:1487–1495.

    https://doi.org/10.1152/jn.1997.77.3.1487

    • PubMed
    • Google Scholar
21. 1. Juusola M
    2. Dau A
    3. Song Z
    4. Solanki N
    5. Rien D
    6. Jaciuch D
    7. Dongre SA
    8. Blanchard F
    9. de Polavieja GG
    10. Hardie RC
    11. Takalo J

    (2017) Microsaccadic sampling of moving image information provides Drosophila hyperacute vision

    eLife 6:e26117.

    https://doi.org/10.7554/eLife.26117

    • PubMed
    • Google Scholar
22. Conference

    1. Kajiya JT

    (1986) The rendering equation the 13th annual conference

    ACM SIGGRAPH Computer Graphics.

    https://doi.org/10.1145/15922.15902

    • Google Scholar
23. Preprint

    1. Kemppainen J
    2. Scales B
    3. Haghighi KR
    4. Takalo J
    5. Mansour N
    6. McManus J
    7. Leko G
    8. Saari P
    9. Hurcomb J
    10. Antohi A
    11. Suuronen JP
    12. Blanchard F
    13. Hardie RC
    14. Song Z
    15. Hampton M
    16. Eckermann M
    17. Westermeier F
    18. Frohn J
    19. Hoekstra H
    20. Lee CH
    21. Huttula M
    22. Mokso R
    23. Juusola M

    (2021) Binocular Mirror-Symmetric Microsaccadic Sampling Enables Drosophila Hyperacute 3D-Vision

    bioRxiv.

    https://doi.org/10.1101/2021.05.03.442473

    • Google Scholar
24. 1. Labhart T
    2. Nilsson DE

    (1995) The dorsal eye of the dragonfly sympetrum: specializations for prey detection against the blue sky

    Journal of Comparative Physiology A 176:437–453.

    https://doi.org/10.1007/BF00196410

    • Google Scholar
25. 1. Lambrinos D
    2. Kobayashi H
    3. Pfeifer R
    4. Maris M
    5. Labhart T
    6. Wehner R

    (1997) An autonomous agent navigating with a polarized light COMPASS

    Adaptive Behavior 6:131–161.

    https://doi.org/10.1177/105971239700600104

    • Google Scholar
26. Book

    1. Land MF

    (1989) Variations in the structure and design of compound eyes

    In: Stavenga DG, Hardie RC, editors. In Facets of Vision. Berlin Heidelberg, Berlin, Heidelberg: Springer. pp. 90–111.

    https://doi.org/10.1007/978-3-642-74082-4_5

    • Google Scholar
27. 1. Land MF
    2. Fernald RD

    (1992) The evolution of eyes

    Annual Review of Neuroscience 15:1–29.

    https://doi.org/10.1146/annurev.ne.15.030192.000245

    • PubMed
    • Google Scholar
28. 1. Land MF

    (1997) Visual acuity in insects

    Annual Review of Entomology 42:147–177.

    https://doi.org/10.1146/annurev.ento.42.1.147

    • PubMed
    • Google Scholar
29. Book

    1. Land MF
    2. Nilsson DE

    (2002)

    Animal Eyes

    Oxford University Press.

    • Google Scholar
30. Conference

    1. Li Z
    2. Liu F
    3. Yang W
    4. Peng S
    5. Zhou J

    (2021) A survey of convolutional neural networks: analysis, applications, and prospects

    IEEE Transactions on Neural Networks and Learning Systems.

    https://doi.org/10.1109/TNNLS.2021.3084827

    • Google Scholar
31. 1. Lippmann RP

    (1987) An introduction to computing with neural nets

    IEEE ASSP Magazine 4:4–22.

    https://doi.org/10.1109/MASSP.1987.1165576

    • Google Scholar
32. 1. Mangan M
    2. Webb B

    (2009) Modelling place memory in crickets

    Biol Cybern 101:307–323.

    https://doi.org/10.1007/s00422-009-0338-1

    • PubMed
    • Google Scholar
33. 1. Meyer EP
    2. Labhart T

    (1993) Morphological specializations of dorsal rim ommatidia in the compound eye of dragonflies and damselfies (Odonata)

    Cell & Tissue Research 272:17–22.

    https://doi.org/10.1007/BF00323566

    • Google Scholar
34. Conference

    1. Millward B
    2. Maddock S
    3. Mangan M
    4. Sheffield Robotics, The University Of Sheffield
    5. Department of Computer Science, The University Of Sheffield

    (2020) Towards Insect Inspired Visual Sensors for Robots

    UKRAS20 Conference.

    https://doi.org/10.31256/Do2Ik3H

    • Google Scholar
35. 1. Möller R

    (2002) Insects could exploit UV-green contrast for landmark navigation

    Journal of Theoretical Biology 214:619–631.

    https://doi.org/10.1006/jtbi.2001.2484

    • PubMed
    • Google Scholar
36. Book

    1. Neumann TR

    (2002) Modeling insect compound eyes: space-variant spherical vision

    In: Poggio TA, editors. Biologically Motivated Computer Vision. Springer. pp. 360–367.

    https://doi.org/10.1007/3-540-36181-2

    • Google Scholar
37. 1. Nityananda V
    2. Bissianna G
    3. Tarawneh G
    4. Read J

    (2016) Small or far away? size and distance perception in the praying mantis

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 371:20150262.

    https://doi.org/10.1098/rstb.2015.0262

    • PubMed
    • Google Scholar
38. 1. Nityananda V
    2. Tarawneh G
    3. Henriksen S
    4. Umeton D
    5. Simmons A
    6. Read JCA

    (2018) A novel form of stereo vision in the praying mantis

    Current Biology 28:588–593.

    https://doi.org/10.1016/j.cub.2018.01.012

    • PubMed
    • Google Scholar
39. 1. Ofstad TA
    2. Zuker CS
    3. Reiser MB

    (2011) Visual place learning in Drosophila melanogaster

    Nature 474:204–207.

    https://doi.org/10.1038/nature10131

    • PubMed
    • Google Scholar
40. 1. Parker SG
    2. Bigler J
    3. Dietrich A
    4. Friedrich H
    5. Hoberock J
    6. Luebke D
    7. McAllister D
    8. McGuire M
    9. Morley K
    10. Robison A
    11. Stich M

    (2010) OptiX: a general purpose ray tracing engine

    ACM Transactions on Graphics 29:1–13.

    https://doi.org/10.1145/1778765.1778803

    • Google Scholar
41. 1. Philippides A
    2. Baddeley B
    3. Cheng K
    4. Graham P

    (2011) How might ants use panoramic views for route navigation?

    The Journal of Experimental Biology 214:445–451.

    https://doi.org/10.1242/jeb.046755

    • PubMed
    • Google Scholar
42. Preprint

    1. Polster J
    2. Petrasch J
    3. Menzel R
    4. Landgraf T

    (2018) Reconstructing the Visual Perception of Honey Bees in Complex 3-D Worlds

    arXiv.

    https://arxiv.org/abs/1811.07560

    • Google Scholar
43. Conference

    1. Purcell TJ
    2. Buck I
    3. Mark WR
    4. Hanrahan P

    (2005) Ray tracing on programmable graphics hardware

    ACM SIGGRAPH 2005 Courses.

    https://doi.org/10.1145/1198555.1198798

    • Google Scholar
44. 1. Risse B
    2. Mangan M
    3. Stürzl W
    4. Webb B

    (2018) Software to convert terrestrial lidar scans of natural environments into photorealistic meshes

    Environmental Modelling & Software 99:88–100.

    https://doi.org/10.1016/j.envsoft.2017.09.018

    • Google Scholar
45. Book

    1. Robinet F
    2. Arnaud R
    3. Parisi T
    4. Cozzi P

    (2014)

    GlTF: designing an open-standard runtime asset format

    In: Engel W, editors. GPU Pro 5: Advanced Rendering Techniques. CRC Press. pp. 18–522.

    • Google Scholar
46. 1. Roth SD

    (1982) Ray casting for modeling solids

    Computer Graphics and Image Processing 18:109–144.

    https://doi.org/10.1016/0146-664X(82)90169-1

    • Google Scholar
47. 1. Scarselli F
    2. Gori M
    3. Tsoi AC
    4. Hagenbuchner M
    5. Monfardini G

    (2009) The graph neural network model

    IEEE Transactions on Neural Networks 20:61–80.

    https://doi.org/10.1109/TNN.2008.2005605

    • PubMed
    • Google Scholar
48. 1. Schwarz S
    2. Narendra A
    3. Zeil J

    (2011) The properties of the visual system in the Australian desert ant melophorus bagoti

    Arthropod Structure & Development 40:128–134.

    https://doi.org/10.1016/j.asd.2010.10.003

    • PubMed
    • Google Scholar
49. Book

    1. Song Z
    2. Coca D
    3. Billings S

    (2009) Biophysical Modeling of a Drosophila Photoreceptor. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

    Springer.

    https://doi.org/10.1007/978-3-642-74082-4_5

    • Google Scholar
50. 1. Song YM
    2. Xie Y
    3. Malyarchuk V
    4. Xiao J
    5. Jung I
    6. Choi KJ
    7. Liu Z
    8. Park H
    9. Lu C
    10. Kim RH
    11. Li R
    12. Crozier KB
    13. Huang Y
    14. Rogers JA

    (2013) Digital cameras with designs inspired by the arthropod eye

    Nature 497:95–99.

    https://doi.org/10.1038/nature12083

    • PubMed
    • Google Scholar
51. Conference

    1. Stone T
    2. Mangan M
    3. Ardin P
    4. Webb B

    (2006) Sky segmentation with ultraviolet images can be used for navigation

    Robotics.

    https://doi.org/10.15607/RSS.2014.X.047

    • Google Scholar
52. 1. Stürzl W
    2. Grixa I
    3. Mair E
    4. Narendra A
    5. Zeil J

    (2015) Three-Dimensional models of natural environments and the mapping of navigational information

    Journal of Comparative Physiology. A, Neuroethology, Sensory, Neural, and Behavioral Physiology 201:563–584.

    https://doi.org/10.1007/s00359-015-1002-y

    • PubMed
    • Google Scholar
53. Conference

    1. Taylor GJ
    2. Baird E

    (2017)

    INSECT VISION: SEGMENTATION TO SIMULATIONS

    In 3rd International Conference on Tomography of Materials and Structures.

    • Google Scholar
54. 1. van Praagh JP
    2. Ribi W
    3. Wehrhahn C
    4. Wittmann D

    (1980) Drone bees fixate the Queen with the dorsal frontal part of their compound eyes

    Journal of Comparative Physiology 136:263–266.

    https://doi.org/10.1007/BF00657542

    • Google Scholar
55. Book

    1. Van Rossum G
    2. Drake FL

    (2009)

    Python 3 Reference Manual

    CreateSpace, Scotts Valley, CA.

    • Google Scholar
56. 1. Vardy A
    2. Moller R

    (2005) Biologically plausible visual homing methods based on optical flow techniques

    Connection Science 17:47–89.

    https://doi.org/10.1080/09540090500140958

    • Google Scholar
57. 1. Viollet S
    2. Franceschini N

    (2010) A hyperacute optical position sensor based on biomimetic retinal micro-scanning

    Sensors and Actuators A 160:60–68.

    https://doi.org/10.1016/j.sna.2010.03.036

    • Google Scholar
58. 1. Viollet S

    (2014) Vibrating makes for better seeing: from the fly’s micro-eye movements to hyperacute visual sensors

    Frontiers in Bioengineering and Biotechnology 2:9.

    https://doi.org/10.3389/fbioe.2014.00009

    • PubMed
    • Google Scholar
59. 1. Wardill TJ
    2. Fabian ST
    3. Pettigrew AC
    4. Stavenga DG
    5. Nordström K
    6. Gonzalez-Bellido PT

    (2017) A novel interception strategy in a miniature robber fly with extreme visual acuity

    Current Biology 27:854–859.

    https://doi.org/10.1016/j.cub.2017.01.050

    • PubMed
    • Google Scholar
60. Book

    1. Wehner R

    (2020) Desert Navigator

    Harvard University Press.

    https://doi.org/10.4159/9780674247918

    • Google Scholar
61. Conference

    1. Whitted T

    (1979) An improved illumination model for shaded display

    the 6th annual conference.

    https://doi.org/10.1145/800249.807419

    • Google Scholar
62. 1. Wiederman SD
    2. Fabian JM
    3. Dunbier JR
    4. O’Carroll DC

    (2017) A predictive focus of gain modulation encodes target trajectories in insect vision

    eLife 6:e26478.

    https://doi.org/10.7554/eLife.26478

    • PubMed
    • Google Scholar
63. 1. Wystrach A
    2. Dewar A
    3. Philippides A
    4. Graham P

    (2016) How do field of view and resolution affect the information content of panoramic scenes for visual navigation? A computational investigation

    Journal of Comparative Physiology. A, Neuroethology, Sensory, Neural, and Behavioral Physiology 202:87–95.

    https://doi.org/10.1007/s00359-015-1052-1

    • PubMed
    • Google Scholar
64. 1. Zeil J
    2. Nalbach G
    3. Nalbach HO

    (1986) Eyes, eye stalks and the visual world of semi-terrestrial crabs

    Journal of Comparative Physiology A 159:801–811.

    https://doi.org/10.1007/BF00603733

    • Google Scholar
65. 1. Zeil J
    2. Hofmann MI
    3. Chahl JS

    (2003) Catchment areas of panoramic snapshots in outdoor scenes

    Journal of the Optical Society of America. A, Optics, Image Science, and Vision 20:450–469.

    https://doi.org/10.1364/josaa.20.000450

    • PubMed
    • Google Scholar

Decision letter after peer review:

Thank you for submitting your article "CompoundRay: An open-source tool for high-speed and high-fidelity rendering of compound eyes" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by myself Albert Cardona as a Reviewing Editor and Tirin Moore as the Senior Editor. The following individuals involved in review of your submission have agreed to reveal their identity: Stephan Saalfeld (Reviewer #1); Andrew D Straw (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

1) All reviewers agree that the raw ommatidial view buffer aspect of the work is the strongest piece and with the most potential to support neuroscience experiments. Expanding on this and discussing the possible applications in simulating insect vision and e.g., interpreting the neural connectomes of insect visual systems is most appropriate. Discussion of the modeling of e.g., insect compound eyes with asymmetric ommatidia is a must; one such example could be robber flies.

2) The motivation for the visualization for human observers is weak, and it is unclear of what use it can be for insect vision or vision research more generally. On surface, it seems a contradiction that the software goes to great lengths to model ommaditia with independent focal points, as necessary for the accurate modeling of each ommatidium, only to then project to a single nodal point to create a camera eye image, contrary to the main point of the work. A stronger motivation with explicit applications that aren't possible otherwise with this camera image are necessary. Otherwise, this aspect of the work ought to be downsized relative to the clearly innovative and useful raw ommaditial view buffer approach for compound eye vision research.

Reviewer #1 (Recommendations for the authors):

More details as they emerged, comments overlap:

98 Do you really mean that each ray has a maximum recursion depth of 0? Did you implement illumination by a light source? If rays do not recurse or trace to light sources, all surfaces are 100% diffuse and ambient and ray casting is equivalent to polygon projection with local illumination. If no illumination by an independent light source is implemented, all surfaces are 100% ambient and ray casting is equivalent to polygon projection. Please clarify.

64 – 67 The first two criteria both cover "arbitrary ommatidia geometry" which I think is one criterion.

65 Spectral sensitivity, which could be criterion 2 is not addressed as far as I understand. The renderer is purely RGB in 8-bit depth, correct? This is not discussed in section "Criterion 2" and could be a big obstacle for real applications. Is it hard to fix?

78 shadow maps are not more resource hungry than ray tracing with direct illumination, can you please explain better or remove?

84 interreflection of light can be implemented by e.g. calculating radiosity maps or doing sampling of rays cast from light sources. I do not see that CompoundRay is doing either, so it is not clear what the relationship to this work is? Can you please clarify?

Overall, I find the paragraph about ray-casting ray-tracing and the concrete relationship with CompoundRay unnecessarily confusing. As said before, I believe it would be better to explain the relevant differences of rendering methods only in how they relate to this project. I.e. how they produce an image (projection rendering by projecting polygons using painters algorithm, texture mapping and/ or colorize with local illumination model; ray-casting by identifying first intersection, texture mapping and/ or colorize with local illumination model, averaging over rays). Mention illumination, shadows (maps for each light in projection, testing for object intersection by tracing to light in ray-tracing) and global models (radiosity or photon sampling) only as possible extensions, not used in this work.

124 – 128 state the above but the language and reasoning are unclear.

227 – 229 eventually clarify that CompoundRay does not currently implement any illumination model or first order ray recursion to cast shadows, so the output would be equivalent to per ommatidium camera rendering and integration, this is way too late in the text as illumination and shading capabilities were used earlier to root for ray-casting for no practical reason.

Reviewer #2 (Recommendations for the authors):

Abstract:

L12-13: I think some expectation of why this might be relevant or interesting would be useful.

L19: Was there any insight that arose from doing this? Especially in regards to the idea raised above – that the shape and overall structure have been ignored?

Intro:

Figure 2: It would be helpful to explicitly indicate which approach is used in the present paper.

Figure 7, top panel and caption for top panel: If I understand it correctly, this top panel is a composite of 100s of individual images but I am confused that I do not see many, if any, vertical edges where one set of samples gives way to another. Also, the jagged effect is not particularly obvious.

Reviewer #3 (Recommendations for the authors):

Below I list a number of small details in need of attention.

Figure 2c could improve, at least by making it larger, or by magnifying a few adjacent ommatidia to better appreciate the effect of not sharing a focal point.

Line 124: Figure 1b is not adequate to illustrate that insect eyes "contain many hundreds or even thousands of lenses".

Line 183: "an image can be generated that captures the full compound eye view." But isn't the point that an image cannot be generated because of the lack of a shared focal point, assuming then that the brain's internal representation of the field of view will be the product of a neural network parsing the output of the many ommatidia? The "cone of influence" of each ommatidium on the final image seems … premature. Neural circuits will combine inputs from ommatidia in ways that have not been worked out in the insect literature, as far as I know. The motivation for wanting to generate an image is unclear: is this for ease of the researchers, which are human, or what is the goal?

The "raw ommatidial view buffer" seems most useful to neuroscientists. The design of a neural network that composes your "Display buffer" from the "raw ommatidial view buffer" would be a most interesting research project, particularly for the opportunity to compare it with the existing, mapped circuit wiring diagram, or connectome, of the Drosophila optic lobes. You could discuss this. In other words, Line 240, "These values are then re-projected into a human-interpretable view" is the potential start of a fascinating research project in its own right, if one redefines this as "fruit-fly interpretable view", as in, the brain of the fruit fly needs to make effective use of this information to successfully navigate the world.

Line 227: "Currently CompoundRay implements only a simple texture lookup, interpolating the nearest pixels of the texture associated with the geometry intersected." What would be gained by, or what is lost now relative to, implementing the full ray tracing approach?

Line 247: "Both pipelines share a common geometry acceleration structure and material shaders so as to save device memory." How is this possible? Can you hand-hold the reader a bit more to clarify how can both pipelines share the same GPU shaders? I may have missed a critical point.

Line 253: "rendering using a traditional camera skips steps 1", you mean, your software can alternatively use traditional cameras in place of modeled ommatidia? I am not sure I get it.

Line 293: "inhomogeneous ommatidial properties". An example you could cite is the eyes of the robber flies, Wardill et al., 2017 Curr Biol., where a frontal fovea with enlarged facets is reported.

Figure 6: seems to imply that overlap is not desirable? Overlap could allow for all kinds of spatial computations, from supraresolution to contrast enhancement to motion detection.

Line 309: the statements would apply only in the absence of a nervous system that would make full use of the ommatidial signals.

And, in general: aren't insect compound eyes known to present significant overlap in the field of view of each ommatidium relative to its neighbors? Figure 6c would then be the most realistic, if ignoring post-processing by the optic lobes?

Figure 7: qualitatively, the top rendered scene looks quite similar throughout. I am not able to notice much difference between the left and right ends. Figure 7 needs work to better convey your message, and to better establish or highlight the vertical relationship between the images in the top and middle and the plot at the bottom.

Line 357: "using a genetic algorithm to find the location", where are the technical details of this approach? Parameters? Framework used, or code? Is it your own code?

Line 440: "On top of the potential biological insights CompoundRay may be able to offer", indeed, your paper would be so much stronger if you showed one such biological insight. It would then become a reference software for compound eye vision.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "CompoundRay, an open-source tool for high-speed and high-fidelity rendering of compound eyes" for further consideration by eLife. Your revised article has been evaluated by Tirin Moore (Senior Editor) and a Reviewing Editor.

The manuscript has been improved but there are some remaining issues that need to be addressed, as outlined below:

Line 234: the sentence starting with "This will aid…" makes claims that, first, aren't entirely clear, and second, seem speculative and belong to the discussion.

Figure 10: the legend is far too short. Needs more detail on what each panel is, particularly for the lower panels. Also, where in the methods is the neural network described? Could be referenced here. Also, "network" is not quite the right shorthand to describe a neural network. Was it a CNN? Or generically an ANN? Or a perceptron?

Line 504: where is the Materials and methods description for the "4-layer fully-connected neural network"?

Line 532: "neural network was trained across 100 epochs three times", why? Why not 200 epochs, or 50, and 2 times, or 10? Why a batch size of 64? Also, these details are hardly worthy of opening the "Experimental results" section. Rather, they belong to the methods, alongside explanations and figures demonstrating why the set of parameters chosen are sensible and appropriate for the problem at hand.

Line 538, isn't the fact that a simple "single eye" design achieved a similar utility score concerning? Perhaps this reveals a disconnect between the test, which runs for 100 epochs (as plotted in Figure 10), and the real world, where an insect needs to take split-second decisions from the let go.

Line 565 reference to "Figure 10 bottom left" could really use the lettering of panels. In addition to a far lengthier and detailed explanation on what, exactly, is in that panel and how it was generated. And if these panels are as critical as they seem to be to evaluate the performance of the 3 eye models, then they ought to be bigger and far more richly annotated, to assist the reader in understanding what is being plotted.

Line 603: if you see opportunities for modeling the "recently reported" (2017) microsaccadic sampling in fruit flies, then why didn't you? Or is this the subject of a paper in preparation or under review elsewhere?

A question for the discussion: with its intrinsically distributed light sampling, compound eyes surely work not with frame-processing downstream circuits but rather event-based circuits, such as those of the dynamic vision sensor or silicon retina (Tobi Delbruck et al., various papers). Could you discuss the implications of asynchronous compound eye-based sensors, particularly when considering uneven ommatidial size which is not available in camera-based eyes such as in mammals or cephalopods? In a fast-moving animal such as a flying honeybee, asynchronous processing surely has advantages for motion detection and obstacle detection and avoidance, and scene recognition (i.e., flowers to land on), or better, to decode the waggle dance language of conspecifics.

Considering event-based computing could make for a much nicer discussion than the caveats of anthropomorphised, frame-based visual computing, or at least complement it well, outlining the intrinsic advantages of compound eyes.

Essential revisions:

1) All reviewers agree that the raw ommatidial view buffer aspect of the work is the strongest piece and with the most potential to support neuroscience experiments. Expanding on this and discussing the possible applications in simulating insect vision and e.g., interpreting the neural connectomes of insect visual systems is most appropriate. Discussion of the modeling of e.g., insect compound eyes with asymmetric ommatidia is a must; one such example could be robber flies.

Thank you for this comment. We have addressed each of the suggestions as follows:

1. To more adequately demonstrate the utility of the raw ommatidial view buffer a new section, “Example Experiment: Apis mellifera Visual Field Comparison” has been added. This offers a more detailed insight into the process of using CompoundRay for a simple experiment analysing the information content available to specific eye designs, posed as a comparative analysis between the arbitrarily-surfaced eyes capable of being supported in CompoundRay and two common approximations of insect visual systems.

2. We have also added an example experiment that demonstrates applications,

3. Introduced a paragraph to the Discussion that outlines immediate areas in which the tool could be used (e.g. asymmetric patterns in robber flies)

2) The motivation for the visualization for human observers is weak, and it is unclear of what use it can be for insect vision or vision research more generally. On surface, it seems a contradiction that the software goes to great lengths to model ommaditia with independent focal points, as necessary for the accurate modeling of each ommatidium, only to then project to a single nodal point to create a camera eye image, contrary to the main point of the work. A stronger motivation with explicit applications that aren't possible otherwise with this camera image are necessary. Otherwise, this aspect of the work ought to be downsized relative to the clearly innovative and useful raw ommaditial view buffer approach for compound eye vision research.

Thank you for the comment. We completely agree with your statement that raw ommatidial view buffer (or, more accurately, a vectorised representation of it) is the more important data generated by CompoundRay. Our intention with the discussion of human visualisation and biassing issues was to highlight, visually, the pitfalls that can occur when attempting to compare two eyes with dissimilar projection surfaces. In addition, we hoped to demonstrate the utility of visualising the insect view in a human-readable way for debugging and sanity-checking purposes, itself with the added caveat that this is susceptible to biases arising from anthropomorphisation as well as projection, and should thus be avoided. To this end, we have made extensive edits to the Methods, Results (including an example experiment), and Discussion which we hope makes this position clearer. Specifically our discussion of visualisation, and the associated cross-comparison issues have been re-framed to focus on the impacts these may have on comparing dissimilar eye designs, making more limited use of the human visualisation scenario as an explanatory example and referencing the work performed in the newly added Example Experiment section.

Reviewer #1 (Recommendations for the authors):

More details as they emerged, comments overlap:

98 Do you really mean that each ray has a maximum recursion depth of 0? Did you implement illumination by a light source? If rays do not recurse or trace to light sources, all surfaces are 100% diffuse and ambient and ray casting is equivalent to polygon projection with local illumination. If no illumination by an independent light source is implemented, all surfaces are 100% ambient and ray casting is equivalent to polygon projection. Please clarify.

Yes, the reviewer is correct that CompoundRay currently uses a recursive depth of 0 (lines 176-179). As a first implementation our aim was to prove that arbitrary surface shapes could be rendered at the speed needed for modelling. More realistic light modelling could be introduced in later advances as outlined in the updated Discussion (line 545-547). Thus, there is an equivalency to polygon projection from each individual ommatidia, which is now discussed and explicitly stated in the paper (lines 136-137), along with reasoning in favour of ray-casting (lines 99-101), namely that it serves as a better platform for implementing new features (As an example, light-source simulation) should they need to be added in the future.

Also see “Changes Relating to Colour and Light Sampling” below.

64 – 67 The first two criteria both cover "arbitrary ommatidia geometry" which I think is one criterion.

These criteria are derived from previously published work, which has been re-worded to be more clearly the case.

65 Spectral sensitivity, which could be criterion 2 is not addressed as far as I understand. The renderer is purely RGB in 8-bit depth, correct? This is not discussed in section "Criterion 2" and could be a big obstacle for real applications. Is it hard to fix?

Yes, the renderer is currently RGB only (Please refer to “Changes Relating to Colour and Light Sampling”) but could be extended to include non-visible light, polarisation etc in the future (See new Discussion). 8-bit has been discussed already in response to point 3.

78 shadow maps are not more resource hungry than ray tracing with direct illumination, can you please explain better or remove?

This section has been removed as suggested.

84 interreflection of light can be implemented by e.g. calculating radiosity maps or doing sampling of rays cast from light sources. I do not see that CompoundRay is doing either, so it is not clear what the relationship to this work is? Can you please clarify?

Overall, I find the paragraph about ray-casting ray-tracing and the concrete relationship with CompoundRay unnecessarily confusing. As said before, I believe it would be better to explain the relevant differences of rendering methods only in how they relate to this project. i.e. how they produce an image (projection rendering by projecting polygons using painters algorithm, texture mapping and/ or colorize with local illumination model; ray-casting by identifying first intersection, texture mapping and/ or colorize with local illumination model, averaging over rays). Mention illumination, shadows (maps for each light in projection, testing for object intersection by tracing to light in ray-tracing) and global models (radiosity or photon sampling) only as possible extensions, not used in this work.

We agree that the previous description of Raycasting was confusing and thus have streamlined the Methods section to reduce confusion on the concrete relationship between ray-casting methods and CompoundRay. We have also made a number of small changes throughout the section entitled “The CompoundRay Software Pipeline” that further clarifies the rendering process. All discussion of advanced light modelling etc has now been downsized as suggested and moved to the Discussion and introduced as future extensions.

Again, more clarification are provided in “Changes Relating to Colour and Light Sampling”

124 – 128 state the above but the language and reasoning are unclear

We have now added reasoning for the of raycasting (lines 99-101) which combined with the newly streamlined Methods, and movement of description of advanced lighting models to the Discussion address the above comment.

227 – 229 eventually clarify that CompoundRay does not currently implement any illumination model or first order ray recursion to cast shadows, so the output would be equivalent to per ommatidium camera rendering and integration, this is way too late in the text as illumination and shading capabilities were used earlier to root for ray-casting for no practical reason.

Explicit description of the equivalency to polygon projection from each individual ommatidia has been added (lines 136-137) as well as explicit statement of recursive depth (lines 176-179). Again, as the focus here is on the geometry modelling of eye models advanced lighting models are considered future work and thus introduced in the updated Discussion.

Please refer to “Changes Relating to Colour and Light Sampling” for additional information.

Reviewer #2 (Recommendations for the authors):

L12-13: I think some expectation of why this might be relevant or interesting would be useful.

Additional explanatory text added (lines 8-9)

L19: Was there any insight that arose from doing this? Especially in regards to the idea raised above – that the shape and overall structure have been ignored?

We have added an example experiment in which we compared realistic vs simple eye designs in the task of small object localisation, which covers this in depth (lines 394-467).

We have also added a sentence to the abstract noting this addition.

Figure 2: It would be helpful to explicitly indicate which approach is used in the present paper.

The approach used in this paper is now included within the figure caption.

Figure 7, top panel and caption for top panel: If I understand it correctly, this top panel is a composite of 100s of individual images but I am confused that I do not see many, if any, vertical edges where one set of samples gives way to another. Also, the jagged effect is not particularly obvious.

This image has now been updated to better communicate the differences between the two views by performing a direct comparison between two points on the sampling curve rather than showing a composite of many points – this makes the difference significantly more visible.

Reviewer #3 (Recommendations for the authors):

Below I list a number of small details in need of attention.

Figure 2c could improve, at least by making it larger, or by magnifying a few adjacent ommatidia to better appreciate the effect of not sharing a focal point.

A magnification inset has been added as suggested.

Line 124: Figure 1b is not adequate to illustrate that insect eyes "contain many hundreds or even thousands of lenses".

The subfigure has been changed to display only one insect eye (that of the Erbenochile), making the density of the lenses on the eye more clear, and including a zoom-section to the diagram of the ommatidium making the linkage between the diagram and the eye image explicit.

Line 183: "an image can be generated that captures the full compound eye view." But isn't the point that an image cannot be generated because of the lack of a shared focal point, assuming then that the brain's internal representation of the field of view will be the product of a neural network parsing the output of the many ommatidia? The "cone of influence" of each ommatidium on the final image seems … premature. Neural circuits will combine inputs from ommatidia in ways that have not been worked out in the insect literature, as far as I know. The motivation for wanting to generate an image is unclear: is this for ease of the researchers, which are human, or what is the goal?

Yes, you are completely correct. Our motivation is to more accurately model the light that arrives to the insect eye ready for analysis and integration into computational models, as such when discussing the “cone of influence” of each ommatidium, we are explicitly describing the light that arrives at the lens. Lines 182-188 have been updated to explicitly state the motivation for creating an image (debugging), and our newly added example experiment (lines 394-467) shows how the realistic eye outputs can be used with artificial neural networks to compare eye designs for specific tasks. This addition exemplifies our intended usage of CompoundRay.

The "raw ommatidial view buffer" seems most useful to neuroscientists. The design of a neural network that composes your "Display buffer" from the "raw ommatidial view buffer" would be a most interesting research project, particularly for the opportunity to compare it with the existing, mapped circuit wiring diagram, or connectome, of the Drosophila optic lobes. You could discuss this. In other words, Line 240, "These values are then re-projected into a human-interpretable view" is the potential start of a fascinating research project in its own right, if one redefines this as "fruit-fly interpretable view", as in, the brain of the fruit fly needs to make effective use of this information to successfully navigate the world.

A new section (outlined below in the “Changes Demonstrating Renderer Use”) has been introduced to give an example of the render’s usage.

Line 227: "Currently CompoundRay implements only a simple texture lookup, interpolating the nearest pixels of the texture associated with the geometry intersected." What would be gained by, or what is lost now relative to, implementing the full ray tracing approach?

The current approach allows arbitrary shaped eyes to be modelled and rendered in a fast way while maintaining the simplicity of the ray-casting approach, which is easier to expand in the future. In essence, compoundRay simulates pinhole-cameras as a regular polygon-projection approach would but doesn’t require abuse of the whole vertex-rendering pipeline to render these multiple cameras. Full ray-tracing could more accurately model light transport through the environments which we introduce as future work in the Discussion. Also see the section below “Changes Relating to Colour and Lighting” below.

Line 247: "Both pipelines share a common geometry acceleration structure and material shaders so as to save device memory." How is this possible? Can you hand-hold the reader a bit more to clarify how can both pipelines share the same GPU shaders? I may have missed a critical point.

This has now been explained (line 264).

Line 253: "rendering using a traditional camera skips steps 1", you mean, your software can alternatively use traditional cameras in place of modeled ommatidia? I am not sure I get it.

Yes, for users to interact with the 3D environment for debugging and analysis purposes CompoundRay allows use of normal cameras as well as ommnitidial-based renderings. We have added a clarification of this point to the text (line 254-255).

Line 293: "inhomogeneous ommatidial properties". An example you could cite is the eyes of the robber flies, Wardill et al. 2017 Curr Biol., where a frontal fovea with enlarged facets is reported.

Thanks for this suggestion. We have added references as requested (lines 298 and 490)

Figure 6: seems to imply that overlap is not desirable? Overlap could allow for all kinds of spatial computations, from supraresolution to contrast enhancement to motion detection.

The caption has been updated to specify that in particular, being able to leverage the most useful parts of homogeneous configurations (i.e. some overlap) is the benefit of a heterogeneous eye design.

Line 309: the statements would apply only in the absence of a nervous system that would make full use of the ommatidial signals.

And, in general: aren't insect compound eyes known to present significant overlap in the field of view of each ommatidium relative to its neighbors? Figure 6c would then be the most realistic, if ignoring post-processing by the optic lobes?

Thank you for this feedback, and you are correct in pointing out the differences between the data shown and real insect eyes. However, Figures5 and 6 are intended not to show accurate compound eye models per se, but rather to show the different information that is provided by different geometries (Figure 5) and varying omnitidial properties (Figure 6). This analysis is intended to convince users of the necessity for this level of detail rather than speak directly to actual insect eyes. However, these are precisely the type of research questions that we have designed CompoundRay to be used to investigate. In particular, figure 6c is modelled as being able to leverage the potentially useful aspects of both the under- and over-sampled views in subfigures a and b, however the exact aspects that are leveraged are left as a simplified selection (just packing cones of vision) to better demonstrate the possibilities present within heterogeneous eye designs.

Figure 7: qualitatively, the top rendered scene looks quite similar throughout. I am not able to notice much difference between the left and right ends. Figure 7 needs work to better convey your message, and to better establish or highlight the vertical relationship between the images in the top and middle and the plot at the bottom.

The image is now a simple split between the 1 and 70-sampled images, with their exact locations on the sampling graph highlighted. This exemplifies the difference between the two and puts the graph into better context.

Line 357: "using a genetic algorithm to find the location", where are the technical details of this approach? Parameters? Framework used, or code? Is it your own code?

We have removed the reference to a genetic algorithm from the text as any optimisation or brute-force search would allow this parameter to be optimised. We feel that our new description better conveys the task to the reader without drastically with specific tools. New text in lines (368-373)

Line 440: "On top of the potential biological insights CompoundRay may be able to offer", indeed, your paper would be so much stronger if you showed one such biological insight. It would then become a reference software for compound eye vision.

Indeed, we have added an entirely new example experiment which shows CompoundRay being used with modern deep-learning tools to investigate small object detection of different eye designs. We feel that this addresses the issue raised by the reviewer and enhances the paper. Thanks for the suggestion. (See the introduced new section outlined below in the “Changes Demonstrating Renderer Use”).

Changes Relating to Colour and Light Sampling

Previous discussion of colour and light sampling was over-complicated and the reasons for differing choices were not justified. We have changed this by removing broad discussion of light transport simulation and instead directly describing what CompoundRay does, particularly within the context of it essentially simulating multiple individual pinhole-cameras in parallel. The reasoning for the selection of ray-casting methods as a basis for the system are mentioned within the methods and discussions sections, and examples of further light and colour sampling methods that could be built on top of CompoundRay are placed in the “Future Extensions” section.

Changes Demonstrating Renderer Use

Following the requests for a specific demonstration of the renderer’s usage we have added an example study that comprises of a comparative analysis of the difference between a more complete model of an insect eye (that includes the surface shape), and two simplified derivatives based on the ideas of pairs of or singular panoramic image sampling methods. This study uses Apis mellifera visual systems as a base, as there already exists a collection of high-fidelity 3D scans of these eyes, one of which was chosen here.

The example study focuses on explaining to the reader the methodology behind using CompoundRay in to perform these comparisons and offers an example path to follow to perform analysis. Basic qualitative analysis is performed on the results, highlighting some surprising behaviours and points of interest.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

The manuscript has been improved but there are some remaining issues that need to be addressed, as outlined below:

Figure 10: the legend is far too short. Needs more detail on what each panel is, particularly for the lower panels. Also, where in the methods is the neural network described? Could be referenced here. Also, "network" is not quite the right shorthand to describe a neural network. Was it a CNN? Or generically an ANN? Or a perceptron?

Line 504: where is the Materials and methods description for the "4-layer fully-connected neural network"?

Line 532: "neural network was trained across 100 epochs three times", why? Why not 200 epochs, or 50, and 2 times, or 10? Why a batch size of 64? Also, these details are hardly worthy of opening the "Experimental results" section. Rather, they belong to the methods, alongside explanations and figures demonstrating why the set of parameters chosen are sensible and appropriate for the problem at hand.

Line 565 reference to "Figure 10 bottom left" could really use the lettering of panels. In addition to a far lengthier and detailed explanation on what, exactly, is in that panel and how it was generated. And if these panels are as critical as they seem to be to evaluate the performance of the 3 eye models, then they ought to be bigger and far more richly annotated, to assist the reader in understanding what is being plotted.

The above comments are addressed together as they relate to the same section and figure. We have taken a number of steps. Firstly, we have separated the previous Figure 10 into 2 figures (Figures5 and 11). Figure 5 describes the experimental procedure only and has a much fuller caption as requested. This supports a new dedicated section in the Methods (Example Inter-Eye Comparison Method, diff pdf lines 280-354) that details the methodology, parameters used etc. Finally, Figure 11 now presents just the example results which links directly to the Results section – we think everything flows better now. Overall, we believe that the new text addresses all of the concerns raised. We appreciate the opportunity to make this clearer for readers.

We have made the main results panels larger and added some annotations to highlight the primary regions of space in which there is a difference in information content for the differing eye designs. However, as outlined above the main contribution of this example experiment is the demonstration of the utility of CompoundRay for modern data-intensive analysis, the pilot data plots are thus intended to inspire users and not define new findings per se.

Line 538, isn't the fact that a simple "single eye" design achieved a similar utility score concerning? Perhaps this reveals a disconnect between the test, which runs for 100 epochs (as plotted in Figure 10), and the real world, where an insect needs to take split-second decisions from the let go.

We did similarly note that there seems little variation in the eye designs and we are sure that this is not related to epoch size. We have added an additional explanation on epoch size selection (diff pdf lines 352-354), and have reframed the task to reflect the use of the neural network (diff pdf lines 280-354) which should clarify this for readers. Specifically, we are learning a close-to-optimal mapping between visual field and position over the course of a number of epochs (here 100), but at test, this mapping is simply used in a single-shot approximation of this mapping function, as is the case with all feedforward neural networks. Thus, the epoch size has no influence at the test time outside of reducing the noise present in the mapping of visual input to relative position.

Line 603: if you see opportunities for modeling the "recently reported" (2017) microsaccadic sampling in fruit flies, then why didn't you? Or is this the subject of a paper in preparation or under review elsewhere?

We have not yet modelled this work, nor is it submitted elsewhere. We see the primary contribution of the paper as a new method that will facilitate multiple new studies of various aspects of compound eyes. We intentionally focussed on the contributions of the methods to keep the focus of the paper clear. This reflects the guidance for Tools and Resources article types: “Tools and Resources articles do not have to report major new biological insights or mechanisms, but it must be clear that they will enable such advances to take place, for example, through exploratory or proof-of-concept experiments. Specifically, submissions will be assessed in terms of their potential to facilitate experiments that address problems that to date have been challenging or even intractable.” We feel that adding more analysis from the pilot study will adversely shift the focus away from Methods and muddy the clarity for the readers.

We note that we are now initiating follow-up work that will use CompoundRay in a far more detailed study of compound eye designs which will be the subject of a separate publication in the future focused on the new insights gained. Inclusion of this work in this manuscript would make it overly long and again deviate from the communication of the Method. Rather we included this text to highlight one of the interesting research topics that this tool could be used to study.

A question for the discussion: with its intrinsically distributed light sampling, compound eyes surely work not with frame-processing downstream circuits but rather event-based circuits, such as those of the dynamic vision sensor or silicon retina (Tobi Delbruck et al., various papers). Could you discuss the implications of asynchronous compound eye-based sensors, particularly when considering uneven ommatidial size which is not available in camera-based eyes such as in mammals or cephalopods? In a fast-moving animal such as a flying honeybee, asynchronous processing surely has advantages for motion detection and obstacle detection and avoidance, and scene recognition (i.e., flowers to land on), or better, to decode the waggle dance language of conspecifics.

Considering event-based computing could make for a much nicer discussion than the caveats of anthropomorphised, frame-based visual computing, or at least complement it well, outlining the intrinsic advantages of compound eyes.

Thank you for pointing out this potential usage of CompoundRay. Indeed, we agree that CompoundRay could be an excellent input for a simulated DVS pipeline due to its speed and reconfigurability. We have now added an extra sentence which makes this potential usage explicit. This potential technological usage compliments well some of the biological research directions mentioned in the previous sentence. (diff pdf lines 626-630)

Author details

1. Blayze Millward

   Department of Computer Science, University of Sheffield, Sheffield, United Kingdom

   Contribution

   Conceptualization, Software, Investigation, Visualization, Methodology, Writing - original draft

   For correspondence

   b.f.millward@sheffield.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0001-9025-1484

2. Steve Maddock

   Department of Computer Science, University of Sheffield, Sheffield, United Kingdom

   Contribution

   Supervision, Validation, Writing - review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-3179-0263

3. Michael Mangan

   Department of Computer Science, University of Sheffield, Sheffield, United Kingdom

   Contribution

   Conceptualization, Supervision, Funding acquisition, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

• 727

  Page views

• 107

  Downloads

• 3

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.