Pith. sign in

REVIEW 3 major objections 6 minor 43 references

EV-LayerSegNet: Self-supervised Motion Segmentation using Event Cameras

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a convolutional network can learn event-based motion segmentation using only a contrast-maximization deblurring loss, with no ground-truth labels or pretrained weights, and that it outperforms two unsupervised…

desk verdict The first fully self-supervised network for event motion segmentation, but 6-of-25 test sequences and no code make the superiority claim unverifiable. read the letter →

arxiv 2506.06596 v1 pith:TDXSEQEY submitted 2025-06-07 cs.CV

Youssef Farah , Federico Paredes-Vallés , Guido De Croon , Muhammad Ahmed Humais , Hussain Sajwani , Yahya Zweiri This is my paper
classification cs.CV
keywords eventcamerasmotionsegmentationself-supervisedlearningcontrastmaximizationaffineopticalflowlayeredrepresentationevent-basedvisionconvolutionalneuralnetwork
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that event-based motion segmentation—deciding which pixels belong to a dominant moving object rather than the background—can be learned without any ground-truth labels or pretrained weights. The proposed network, EV-LayerSegNet, estimates two affine motion models and two soft segmentation masks from a batch of raw events, combines them into one flow field, and warps the events to a reference time. The only training signal is contrast maximization: the sharper the image of warped events, the more accurate the flow and masks are assumed to be. On a simulated dataset restricted to affine motion, the method reaches up to 71% IoU and 87% detection rate, outperforming the two unsupervised baselines on most reported sequences. The significance is that expensive manual or sensor-based annotation, a major bottleneck for event cameras, may be replaceable by a loss derived from the events themselves.

What carries the argument

The load-bearing object is the pair of differentiable segmentation masks $\alpha_1, \alpha_2$ from the segmentation module, paired with the two affine motion fields $W_1, W_2$ from the optical flow module. Masks and flows are combined as $W_{\mathrm{comb}} = \alpha_1 \odot W_1 + \alpha_2 \odot W_2$, and events are warped according to $W_{\mathrm{comb}}$; the squared average-timestamp image of warped events, summed over both polarities and over forward and backward warping, is the sole training loss. A softmax-plus-maxout operation keeps pixel-to-layer assignment differentiable while ensuring each pixel belongs to one layer. The leaky DoReLU activation, whose slope is nearly flat outside $[0,1]$, is shown by the ablation to be important for segmentation quality.

What would settle it

Compute the contrast loss and the mask IoU for every test sequence and check whether the sequences with the lowest loss also have the highest IoU; if a degenerate solution can be produced—for example, masks that split the scene into arbitrary regions while the two affine flows still align most events—that reaches near-zero loss with near-zero IoU, then the self-supervision signal is insufficient. A more direct experiment would freeze the optical flow module, optimize only the masks under $L_{\mathrm{contrast}}$, and test whether true object boundaries emerge.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a two-layer decomposition of an event stream is learnable end-to-end with no supervision beyond the event data. The network outputs an alpha mask for each of two layers and six affine motion parameters per layer; the masks are made differentiable by a softmax-plus-maxout operation, so each pixel belongs to exactly one layer while gradients still flow. The two flows are fused into a single combined flow, events are warped forward and backward, and the squared average timestamp of the warped events serves as the loss. Because events produced by the same edge should collapse onto the edge when the flow is correct, the paper argues that minimizing this deblurring loss drives the masks to separate the two motions. The empirical claim is that this works on simulated affine scenes and beats the iterative unsupervised baselines on most of the six reported test sequences, with inference in milliseconds rather than minutes.

Load-bearing premise

The load-bearing premise is that the single deblurring loss, contrast maximization on the combined affine flow, is enough to push the segmentation masks toward true object boundaries; the paper offers no proof or diagnostic that a low loss implies correct masks, so a degenerate or speckled mask assignment that still deblurs the events could in principle satisfy the loss.

Editorial extensions

If this is right

  • Motion segmentation for event cameras can be trained on raw event streams alone, removing the need for per-event ground truth that is expensive and difficult to obtain at microsecond resolution.
  • On the reported simulated test sequences, the method reaches up to 71% IoU and 87% detection rate, outperforming the iterative unsupervised baselines on most sequences while running in milliseconds rather than minutes.
  • The ablations indicate that the near-flat slope of leaky DoReLU outside $[0,1]$ matters: replacing it with leaky ReLU or a steeper slope ($\gamma=10$) substantially lowers segmentation performance.
  • The current formulation is restricted to affine motion and is sensitive to very high blur and noise, so the demonstrated gains are for simulated affine scenes; extending the motion model is the stated path toward real-world 3D data.
  • If the method generalizes beyond the simulator, it offers a route to onboard motion segmentation for drones and other weight-constrained platforms, where iterative optimization baselines are too slow.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, the load-bearing assumption could be tested by freezing the optical flow module, training the masks alone under the same contrast loss, and checking whether object boundaries emerge.
  • Beyond the paper, the deblurring loss itself could serve as a confidence score, so the network would only trust its segmentation when the loss is low, making the method usable with no labels at deployment.
  • Beyond the paper, the two-layer affine assumption could be relaxed to per-layer non-rigid or spline motion models, and the same loss would still provide supervision at the cost of more parameters.
  • Beyond the paper, combining the contrast loss with photometric constancy or egomotion constraints could yield joint estimation of camera motion and independently moving objects, a step toward real scenes that violate affine motion.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes EV-LayerSegNet, a CNN for event-based motion segmentation trained in a self-supervised manner via contrast maximization. The network estimates two affine motion models and two soft alpha masks, combines them into a single optical flow map, warps the input events, and uses the deblurring quality of the warped events as its only training loss. Experiments on a simulated ESIM dataset are compared against EMSMC and EMSGC, with IoU and detection rate reported on six selected test sequences; ablations study the Leaky DoReLU activation and its slope parameter. The central claim is that this self-supervised method outperforms prior unsupervised event-based motion segmentation methods.

Significance. If the claimed results were representative of the full test set, this would be a notable contribution: it would demonstrate a learning-based event-camera motion segmentation method that requires no ground truth annotations and no pretrained weights, with fast inference and a clear two-layer affine motion model. The training signal (contrast maximization) is genuinely distinct from the evaluation metrics (IoU and detection rate), so there is no direct metric-fitting circularity. The paper also contributes a new simulated dataset and ablates an architectural choice (Leaky DoReLU vs Leaky ReLU). However, the empirical evidence is not yet sufficient to support the stated superiority over state-of-the-art methods.

major comments (3)
  1. [Section 6.2, Table 1] The quantitative evaluation reports only 6 of the 25 test sequences defined in Section 5.2. The sentence "All methods have found challenging the test dataset, having relevant performances in 6 sequences" indicates that the remaining 19 sequences were excluded, but no criterion for "relevant" is given, no full-dataset means or variances are reported, and no per-sequence results cover failures. Since Contribution 3 claims superiority over state-of-the-art unsupervised methods, this selective reporting makes the central empirical claim unsupported. Please report results for all 25 sequences, including failures, with aggregate statistics and a per-sequence table.
  2. [Section 3.3, Eq. (8)] The contrast-maximization loss is the sole training objective, but the paper provides no evidence that a low deblurring loss implies correct segmentation masks. A degenerate solution could assign arbitrary or speckled alpha maps while still producing a deblurred event image, for example if one layer absorbs most events or if the two affine flow maps agree on the dominant motion. To support the statement that deblurring quality consequently means the masks are accurate, the authors should add a diagnostic (e.g., a plot of training loss against validation IoU, or an ablation with randomized/frozen masks) or provide an identifiability argument for the two-layer affine model under this loss.
  3. [Section 6.2, Table 1, row "Bird in front of building"] On this sequence the proposed method is substantially worse than EMSGC (IoU 0.10 vs 0.41, DR 0.00 vs 0.21). This failure case is directly relevant to the claim of superiority and should be discussed. The sequence is also omitted from the ablation tables (Tables 2 and 3 list only five sequences), so the reader cannot see whether the design choices that help on the other sequences also explain this failure. Please include this case in the ablations or explain why it is excluded.
minor comments (6)
  1. [Section 3.3] There is a typo: "misaglined" should be "misaligned", and "the the per-pixel" in the loss description should be "the per-pixel".
  2. [Section 6.2, Table 1] The caption says "using an identity camera matrix", but the relevance of the camera matrix to the comparison is not explained; please clarify what this means for the simulation and the evaluation.
  3. [Eq. (13)] The detection rate definition is unclear: the condition "BD ∩ BG > 0.5" appears to use an unstated threshold, and the second condition "(BD ∩ BG) > (SD ∩ BG)" is not dimensionally meaningful as written. Please define the areas and thresholds precisely.
  4. [Section 4.1] The optical flow module is described as ending with a feedforward network of 4 layers, but the final layer activation is not specified; please state whether the last layer is linear or tanh.
  5. [Section 3.2] The "maxout operation" is described as making the one-hot classification differentiable; it would be more accurate to say it is differentiable almost everywhere, since the hard assignment itself is non-differentiable at ties. This is a minor wording issue but could confuse readers.
  6. [Section 5.2] The ground truth is generated "by using a threshold on optical flow" because foreground objects move faster; the actual threshold value is never reported. Please provide the threshold and justify that it yields correct object labels in all 25 test sequences.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the self-supervised loss and the reported evaluation metrics are distinct, and no load-bearing claim reduces to a fitted input or a self-citation chain.

full rationale

The paper's derivation chain is: input event volume -> jointly estimated segmentation masks and affine flow maps -> combined flow Wcomb = alpha1*W1 + alpha2*W2 -> event warping -> contrast maximization loss (Eqs. 7-10). The training signal is a deblurring/contrast objective computed from the combined flow, while the reported IoU and detection rate are computed against ground-truth masks derived by thresholding the simulated optical flow (Section 5.2). The ground truth is never used during training, and no fitted parameter is renamed as a prediction. The central assumption that better deblurring implies better segmentation is an empirical hypothesis, not a circular definition: the paper evaluates this hypothesis on held-out sequences and the metric is independent of the loss. Citations to prior work, including [9] whose authors overlap with the present paper, provide standard, independently published components (contrast maximization loss, encoder-decoder layout, leaky DoReLU activation) that are reimplemented and benchmarked here; they are not invoked as unverified uniqueness theorems or as the sole justification of the central claim. The selective reporting of 6 of 25 test sequences in Table 1 is a legitimate evidence-quality concern, but it concerns the support for the empirical claim, not circularity of the derivation. No specific reduction of the claimed result to its own inputs was found.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a two-layer affine motion model and the empirical hypothesis that contrast maximization provides a sufficient self-supervised signal for segmentation. These are domain assumptions, not standard mathematical axioms, and they are not validated independently. No new physical entities are introduced.

free parameters (3)
  • smoothness weight lambda = 0.001
    Hand-chosen weight balancing the Charbonnier smoothness prior in Eq. 10; influences flow regularity but is not fitted to data.
  • leaky DoReLU slope gamma = 100
    Hand-chosen slope for the segmentation activation in Eq. 11; the ablation (gamma=10) shows performance degrades significantly, so this choice materially affects results.
  • input event count N = 200,000 for training, 50,000 for all methods at test
    Chosen to ensure enough motion blur in the input partition; the test value is also used for baselines to keep the comparison feasible.
assumptions (5)
  • domain assumption Scene motion is affine and can be modelled as two layers, each with a six-parameter affine motion matrix (Eq. 4).
    This restricts the method to affine motions, acknowledged in the conclusion. It is essential for the flow parameterization and the two-layer segmentation.
  • domain assumption Constant illumination, so events are generated only by relative motion between camera and scene.
    Assumed in Section 3.2 to justify that deblurring quality measures motion correctness; under varying illumination the loss could be distorted.
  • domain assumption Contrast maximization (minimizing the average timestamp of warped events) is a valid proxy for correct optical flow and segmentation.
    Section 3.3 uses L_contrast (Eq. 8) as the sole supervision signal. If a low deblurring loss could be achieved with incorrect masks, the method would not learn the intended segmentation.
  • domain assumption Test ground truth can be obtained by thresholding optical flow because foreground objects move significantly faster than the background.
    Section 5.2; this threshold-based GT is used to compute IoU and detection rate, so evaluation accuracy depends on this assumption holding in the synthetic scenes.
  • standard math Softmax plus maxout yields differentiable masks that sum to one and assign each pixel to a single layer.
    Section 3.2; standard differentiable relaxation to handle discrete allocation, a common technique in layered segmentation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EV-LayerSegNet: Self-supervised Motion Segmentation using Event Cameras." pith.science (2026). https://pith.science/paper/TDXSEQEY

@misc{pith2026250606596,
  author       = {Pith},
  title        = {Pith review of: EV-LayerSegNet: Self-supervised Motion Segmentation using Event Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TDXSEQEY}},
  note         = {Machine review of arXiv:2506.06596}
}
read the original abstract

Event cameras are novel bio-inspired sensors that capture motion dynamics with much higher temporal resolution than traditional cameras, since pixels react asynchronously to brightness changes. They are therefore better suited for tasks involving motion such as motion segmentation. However, training event-based networks still represents a difficult challenge, as obtaining ground truth is very expensive, error-prone and limited in frequency. In this article, we introduce EV-LayerSegNet, a self-supervised CNN for event-based motion segmentation. Inspired by a layered representation of the scene dynamics, we show that it is possible to learn affine optical flow and segmentation masks separately, and use them to deblur the input events. The deblurring quality is then measured and used as self-supervised learning loss. We train and test the network on a simulated dataset with only affine motion, achieving IoU and detection rate up to 71% and 87% respectively.

Figures

Figures reproduced from arXiv: 2506.06596 by the authors.

Figure 1
Figure 1. Our method takes as input an event volume that results in a blurry scene. It then attempts to generate two different masks to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. EV-LayerSegNet architecture. Events are downsampled by 4 encoding layers before being passed to 2 residual blocks. The output [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Test result example on synthetic dataset (”Bird in front of building” in Tab. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visualization of leaky ReLU (left) and leaky DoReLU [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 4
Figure 4. Figure 4: Qualitative comparison of test results from [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Neuromorphic vision-based motion segmentation with graph transformer neural network.Trans

    Yusra Alkendi, Rana Azzam, Sajid Javed, Lakmal Senevi- ratne, and Yahya Zweiri. Neuromorphic vision-based motion segmentation with graph transformer neural network.Trans. Multi., 27:385–400, 2025. 4

    work page 2025

  2. [2]

    Motion segmentation for neu- romorphic aerial surveillance, 2024

    Sami Arja, Alexandre Marcireau, Saeed Afshar, Bharath Ramesh, and Gregory Cohen. Motion segmentation for neu- romorphic aerial surveillance, 2024. 4

    work page 2024

  3. [3]

    Charbonnier, L

    P. Charbonnier, L. Blanc-Feraud, G. Aubert, and M. Bar- laud. Two deterministic half-quadratic regularization algo- rithms for computed imaging. InProceedings of 1st Inter- national Conference on Image Processing, pages 168–172 vol.2, 1994. 5

    work page 1994

  4. [4]

    Focus is all you need: Loss functions for event-based vision

    Guillermo Gallego, Mathias Gehrig, and Davide Scara- muzza. Focus is all you need: Loss functions for event-based vision. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12272–12281, 2019. 5

    work page 2019

  5. [5]

    Davison, Jorg Conradt, Kostas Daniilidis, and Davide Scaramuzza

    Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jorg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey.IEEE Transactions on Pattern Analysis & Machine Intelligence, 44 (01):154–180, 2022. 1, 2

    work page 2022

  6. [6]

    Opti- cal flow estimation from layered nearest neighbor flow fields

    Liyue Ge, Congxuan Zhang, Zhen Chen, and Ming Li. Opti- cal flow estimation from layered nearest neighbor flow fields. In2018 11th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP- BMEI), pages 1–6, 2018. 2

    work page 2018

  7. [7]

    Dsec: A stereo event camera dataset for driv- ing scenarios.IEEE Robotics and Automation Letters, 6(3): 4947–4954, 2021

    Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driv- ing scenarios.IEEE Robotics and Automation Letters, 6(3): 4947–4954, 2021. 2

    work page 2021

  8. [8]

    Out of the Room: Generalizing Event-Based Dynamic Motion Seg- mentation for Complex Scenes

    Stamatios Georgoulis, Weining Ren, Alfredo Bochicchio, Daniel Eckert, Yuanyou Li, and Abel Gawel. Out of the Room: Generalizing Event-Based Dynamic Motion Seg- mentation for Complex Scenes . In2024 International Con- ference on 3D Vision (3DV), pages 442–452, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 4

    work page 2024

Show all 43 references
  1. [9]

    Self-supervised learning of event-based optical flow with spiking neural networks

    Jesse Hagenaars, Federico Paredes-Valles, and Guido de Croon. Self-supervised learning of event-based optical flow with spiking neural networks. InAdvances in Neural Infor- mation Processing Systems, pages 7167–7179. Curran Asso- ciates, Inc., 2021. 4, 5, 7

    2021

  2. [10]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980, 2014. 6

    2014 arXiv

  3. [11]

    Kumar, Philip Torr, and A

    M. Kumar, Philip Torr, and A. Zisserman. Learning lay- ered motion segmentations of video.International Journal of Computer Vision, 76:301–319, 2008. 2

    2008

  4. [12]

    Spike-flownet: Event-based optical flow estimation with energy-efficient hy- brid neural networks

    Chankyu Lee, Adarsh Kosta, Alex Zihao Zhu, Kenneth Chaney, Kostas Daniilidis, and Kaushik Roy. Spike-flownet: Event-based optical flow estimation with energy-efficient hy- brid neural networks. InEuropean Conference on Computer Vision, 2020. 4

    2020

  5. [13]

    Fusion-flownet: Energy-efficient optical flow estimation us- ing sensor fusion and deep fused spiking-analog network ar- chitectures

    Chankyu Lee, Adarsh Kumar Kosta, and Kaushik Roy. Fusion-flownet: Energy-efficient optical flow estimation us- ing sensor fusion and deep fused spiking-analog network ar- chitectures. In2022 International Conference on Robotics and Automation (ICRA), pages 6504–6510, 2022. 4

    2022

  6. [14]

    Siyang Li, Bryan Seybold, Alexey V orobyov, Alireza Fathi, Qin Huang, and C.-C. Jay Kuo. Instance embedding trans- fer to unsupervised video object segmentation. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6526–6535, 2018. 3

    2018

  7. [15]

    A lightweight net- work to learn optical flow from event data

    Zhuoyan Li, Jiawei Shen, and Ruitao Liu. A lightweight net- work to learn optical flow from event data. In2020 25th International Conference on Pattern Recognition (ICPR), pages 1–7, 2021. 4

    2021

  8. [16]

    Belongie, Lubomir D

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll’a r, and C. Lawrence Zitnick. Microsoft COCO: common objects in context.CoRR, abs/1405.0312,

    arXiv

  9. [17]

    See more, know more: Unsuper- vised video object segmentation with co-attention siamese networks

    Xiankai Lu, Wenguan Wang, Chao Ma, Jianbing Shen, Ling Shao, and Fatih Porikli. See more, know more: Unsuper- vised video object segmentation with co-attention siamese networks. In2019 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 3618–3627,

  10. [18]

    Event-based moving object detection and tracking

    Anton Mitrokhin, Cornelia Ferm ¨uller, Chethan Paramesh- wara, and Yiannis Aloimonos. Event-based moving object detection and tracking. In2018 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), pages 1–9,

  11. [19]

    Learning visual motion segmentation using event surfaces

    Anton Mitrokhin, Zhiyuan Hua, Cornelia Ferm ¨uller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14402–14411, 2020. 2, 3

    2020

  12. [20]

    Parameshwara, Simin Li, Cornelia Ferm ¨uller, Nitin J

    Chethan M. Parameshwara, Simin Li, Cornelia Ferm ¨uller, Nitin J. Sanket, Matthew S. Evanusa, and Yiannis Aloi- monos. Spikems: Deep spiking neural network for motion segmentation. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3414–3420,

  13. [21]

    Parameshwara, Nitin J

    Chethan M. Parameshwara, Nitin J. Sanket, Chahat Deep Singh, Cornelia Ferm ¨uller, and Yiannis Aloimonos. 0- mms: Zero-shot multi-motion segmentation with a monocu- lar event camera. In2021 IEEE International Conference on Robotics and Automation (ICRA), pages 9594–9600, 2021. 2, 3

    2021

  14. [22]

    Federico Paredes-Vall ´es and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image recon- struction for event cameras via photometric constancy. In 2021 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 3445–3454, 2021. 4

    2021

  15. [23]

    Learning features by watching ob- jects move

    Deepak Pathak, Ross Girshick, Piotr Doll ´ar, Trevor Darrell, and Bharath Hariharan. Learning features by watching ob- jects move. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6024–6033, 2017. 3

    2017

  16. [24]

    ESIM: an open event camera simulator.Conf

    Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator.Conf. on Robotics Learning (CoRL), 2018. 6

    2018

  17. [25]

    Sanket, Chethan M

    Nitin J. Sanket, Chethan M. Parameshwara, Chahat Deep Singh, Ashwin V . Kuruttukulam, Cornelia Ferm¨uller, Davide Scaramuzza, and Yiannis Aloimonos. Evdodgenet: Deep dynamic obstacle dodging with event cameras. In2020 IEEE International Conference on Robotics and Automation (I...

    2020

  18. [26]

    Learning to segment dominant object motion from watching videos

    Sahir Shrestha, Mohammad Ali Armin, Hongdong Li, and Nick Barnes. Learning to segment dominant object motion from watching videos. In2021 Digital Image Computing: Techniques and Applications (DICTA), pages 01–08, 2021. 2, 3, 4, 5, 6, 7

    2021

  19. [27]

    Event-based motion segmentation by motion compensation

    Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7243–7252, 2019. 2, 3, 6, 7, 8

    2019

  20. [28]

    Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert E. Mahony. Reducing the sim-to-real gap for event cameras. InEuropean Conference on Computer Vision,

  21. [29]

    Wang and E.H

    J.Y .A. Wang and E.H. Adelson. Representing moving im- ages with layers.IEEE Transactions on Image Processing, 3 (5):625–638, 1994. 2

    1994

  22. [30]

    Un- evmoseg: Unsupervised event-based independent motion segmentation, 2023

    Ziyun Wang, Jinyuan Guo, and Kostas Daniilidis. Un- evmoseg: Unsupervised event-based independent motion segmentation, 2023. 4

    2023

  23. [31]

    Self-supervised video object segmentation by motion grouping

    Charig Yang, Hala Lamdouar, Erika Lu, Andrew Zisserman, and Weidi Xie. Self-supervised video object segmentation by motion grouping. In2021 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pages 7157–7168, 2021. 3

    2021

  24. [32]

    Multi-motion and ap- pearance self-supervised moving object detection

    Fan Yang, Srikrishna Karanam, Meng Zheng, Terrence Chen, Haibin Ling, and Ziyan Wu. Multi-motion and ap- pearance self-supervised moving object detection. In2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2101–2110, 2022. 3

    2022

  25. [33]

    Unsupervised moving object detection via contextual information separation

    Yanchao Yang, Antonio Loquercio, Davide Scaramuzza, and Stefano Soatto. Unsupervised moving object detection via contextual information separation. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 879–888, 2019. 3

    2019

  26. [34]

    Yorke, and Yiannis Aloimonos

    Chengxi Ye, Anton Mitrokhin, Cornelia Ferm ¨uller, James A. Yorke, and Yiannis Aloimonos. Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5831–5838,

  27. [35]

    Deformable sprites for unsupervised video decomposition

    Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2657–2666, 2022. 3

    2022

  28. [36]

    Bdd100k: A diverse driving video database with scalable annotation tool- ing.ArXiv, abs/1805.04687, 2018

    Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving video database with scalable annotation tool- ing.ArXiv, abs/1805.04687, 2018. 2

    2018 arXiv

  29. [37]

    Enhanced local subspace affinity for feature- based motion segmentation.Pattern Recognition, 44:454– 470, 2011

    Luca Zappella, Xavier Llado, Edoardo Provenzi, and Joaquim Salvi. Enhanced local subspace affinity for feature- based motion segmentation.Pattern Recognition, 44:454– 470, 2011. 2

    2011

  30. [38]

    Event-enhanced snapshot compressive videography at 10k fps.IEEE Trans

    Bo Zhang, Jinli Suo, and Qionghai Dai. Event-enhanced snapshot compressive videography at 10k fps.IEEE Trans. Pattern Anal. Mach. Intell., 47(2):1266–1278, 2025. 1

    2025

  31. [39]

    Permutation pref- erence based alternate sampling and clustering for motion segmentation.IEEE Signal Processing Letters, 25(3):432– 436, 2018

    Yun Zhang, Bin Luo, and Liangpei Zhang. Permutation pref- erence based alternate sampling and clustering for motion segmentation.IEEE Signal Processing Letters, 25(3):432– 436, 2018. 2

    2018

  32. [40]

    Event-based motion segmentation with spatio- temporal graph cuts.IEEE Transactions on Neural Networks and Learning Systems, pages 1–13, 2021

    Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatio- temporal graph cuts.IEEE Transactions on Neural Networks and Learning Systems, pages 1–13, 2021. 2, 3, 6, 7, 8

    2021

  33. [41]

    The multi- vehicle stereo event camera dataset: An event camera dataset for 3d perception.IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018

    Alex Zihao Zhu, Dinesh Thakur, Tolga ¨Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi- vehicle stereo event camera dataset: An event camera dataset for 3d perception.IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 2

    2018

  34. [42]

    Ev-flownet: Self-supervised optical flow estimation for event-based cameras.ArXiv, abs/1802.06898,

    Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Ev-flownet: Self-supervised optical flow estimation for event-based cameras.ArXiv, abs/1802.06898,

    arXiv

  35. [43]

    Unsupervised event-based learning of op- tical flow, depth, and egomotion

    Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of op- tical flow, depth, and egomotion. In2019 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 989–997, 2019. 2, 4, 5

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.