Motion-aware Event Suppression for Event Cameras

Davide Scaramuzza; Giovanni Cioffi; Marco Cannici; Nico Messikommer; Roberto Pellerito

REVIEW 4 major objections 4 minor 35 references

Motion-aware Event Suppression for Event Cameras

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that event streams from a moving camera can be decoupled in real time by jointly segmenting independently moving objects and forecasting their future motion, enabling anticipatory suppression of dynamic events before they

desk verdict Solid, well-engineered new task (anticipatory event suppression) with overstated headline numbers; the core method holds, but the formal support for 'before they occur' is weaker than the abstract suggests. read the letter →

arxiv 2602.23204 v4 pith:MRXIWQJ4 submitted 2026-02-26 cs.CV cs.RO

Roberto Pellerito , Nico Messikommer , Giovanni Cioffi , Marco Cannici , Davide Scaramuzza This is my paper

classification cs.CVcs.RO

keywords eventcamerasmotionsegmentationopticalflowforecastingsuppressionattention-basedtimeconditioningvisualodometrytokenpruninganticipatoryprediction

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 introduces a learning-based method for event cameras that separates events caused by the camera's own motion from events caused by independently moving objects (IMOs). The central idea is to not only locate IMOs in the current event stream but also predict where they will be over the next 100 ms, so that the wrong events can be suppressed before they arrive. The authors argue that this anticipatory filtering has been missing: prior work either computed dense 3D reconstructions, which is too slow, or used hand-tuned biological filters, which are brittle. If the method works as claimed, downstream perception systems such as visual odometry, object segmentation, and AR/VR can run on a much cleaner, lower-latency event stream, with the reported 173 Hz throughput on a consumer GPU.

What carries the argument

The central machinery is the motion-suppression operator defined on a spatio-temporal event window, made tractable by assuming it is time-independent within the window. The network realizes it with three components: a conv-GRU encoder that turns a 50 ms event slice into spatial features; an attention-based time conditioning (ATC) module that cross-attends a sinusoidal temporal encoding of the target forecasting time against the spatial features, allowing one network to produce flow for arbitrary horizons; and a differentiable backward-warping step that propagates soft mask logits along the predicted flow. The ATC module is what carries the forecasting ability: it converts the scalar predicti

What would settle it

Take a high-speed sequence where an independently moving object crosses more than one pixel per millisecond and disappears behind an occlusion inside the 100 ms horizon; if the warped mask's IoU with the true future object falls below what a static, no-forecast mask achieves, then the anticipatory suppression claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that motion suppression can be decomposed into two jointly learned tasks—binary segmentation of independently moving objects and dense optical flow forecasting for a short horizon—and that composing them through a learned mask-warping step yields zero-latency suppression. Specifically, it proposes a recurrent encoder that outputs both a current IMO mask and a future flow field, guided by an attention-based time conditioning module that modulates spatial features with a continuous temporal encoding of the target horizon. Warping the mask forward with the predicted flow produces an anticipated mask that is applied to gate incoming events. On the EVIMO benchmark, th

Load-bearing premise

The whole gating rule is a single mask at one reference time, warped by one predicted flow field and held constant over the window; if an object changes shape or two objects overlap inside that window, the mask mislabels events.

Editorial extensions

If this is right

If the central claim holds, event cameras can deliver a causally filtered stream with effectively zero-latency masks, since the 100 ms forecast covers the 5.8 ms inference cost.
A cleaner ego-motion-only stream should improve any downstream estimator that assumes static scenes: the paper reports an 8.7% ATE reduction in its visual odometry experiment, approaching the ground-truth upper bound.
Motion-guided token pruning in vision transformers becomes practical: the paper reports an 83% acceleration with under 7 points of segmentation AP drop on the driving dataset.
The same mask-plus-flow forecast can be recomputed at 173 Hz, meaning the suppressor can be inserted as a preprocessing stage without becoming the bottleneck.
The horizon ablation suggests a tunable latency-accuracy trade-off: selecting shorter prediction horizons in fast-moving scenes preserves accuracy while still providing anticipation.

Reading between the lines

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

I infer that the time-conditioning design could be reused beyond suppression: any event-based predictor that needs to answer "where will things be in τ milliseconds" — tracking, planning, or reactive control — could adopt the same cross-attention conditioning mechanism.
Since the unsupervised contrast-maximization loss provides dense flow supervision without labels, extending the method to new environments may only require unlabeled event streams, making it a plausible self-supervised preprocessing layer.
The paper's time-independence assumption means a single flow field must explain all IMO motion inside the window; in scenes with multiple objects moving at different depths and velocities, the mask warp may smear boundaries. A testable extension is to predict a small set of flow modes or per-pixel uncertainty alongside the flow.
The monotonic mIoU decay with horizon implies that an adaptive controller could choose the prediction horizon based on scene dynamics estimated from recent flow magnitude, keeping accuracy high while preserving anticipation.

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.

Desk Editor's Note

read the letter

Two things to know. First, this is a genuinely new task formulation: jointly segment IMOs in the current event stream while predicting their future motion, then warp the mask forward to gate events. That combination is not in the cited prior work, and the engineering is competent—conv-GRU encoder, an attention-based time-conditioning module that modulates features by target horizon, and differentiable mask warping. Second, the headline numbers overstate what the experiments show, and there is a real gap between the formal claim of suppressing future events and what is actually evaluated.

Credit where due: the method runs at 173 Hz on a 2080 Ti with under 1 GB memory, and it beats existing baselines on EVIMO future segmentation and current pIoU. The downstream demonstrations—VO improvement and ViT token pruning—are useful sanity checks rather than filler. They also retrain baseline EVIMO and OMS variants, including a strict causal baseline, which is more careful than most papers in this area. The horizon ablation in Fig. 6 is honest: mIoU drops monotonically as the prediction horizon grows, and fast sequences degrade faster.

The main soft spot is the mismatch between the math and the claim. Equation (3) defines the suppression operator on the past window E[t−Δt,t). Equation (3*) warps a lagging mask to the current time. But the pipeline description and Fig. 2 claim to suppress future events E[t, t+Δtp). The formal operator never actually gates future events; it gates past events with a warped mask. The time-independence assumption in Section III-A—"it suffices to determine the spatial support once"—is exactly the single-warp approximation that the stress-test flags, and the paper never evaluates event-level correctness inside the suppression window, only mask IoU at the horizon endpoint. So the anticipatory suppression is real as a forecast, but the formal grounding for "before they occur" is not fully established.

Also: the abstract and conclusion claim a 13% ATE reduction, but Section IV-G reports −8.7% and says 13% is the ground-truth-mask upper bound. The 67% figure is R@0.5 against the strict causal EVIMO†; against the standard EVIMO∗ it's 45%—still good, but the abstract doesn't say. No error bars on the main tables, and no code is provided. The conclusion says "operating 60 Hz faster" while the abstract says 53% higher inference rate; minor, but sloppy.

None of this is fatal. The architecture is sensible, the ablations show ATC and flow warping help, and the qualitative results look convincing. The task is worth building on. This paper should go to peer review, not be desk-rejected. A serious referee should ask for a corrected abstract/conclusion, a precise formal statement of what the suppression operator acts on, and an event-level evaluation of suppression quality inside the prediction window—or at minimum an explicit statement that the current evaluation measures mask forecast accuracy only.

Referee Report

4 major / 4 minor

Summary. The paper proposes Motion-aware Event Suppression: a recurrent encoder-decoder with an attention-based time conditioning (ATC) module that jointly predicts a binary IMO mask and a future dense flow, warps the mask forward, and uses it to gate events. Training combines supervised BCE/Dice on current and future masks, supervised/unsupervised flow losses, and contrast maximization. Evaluations on EVIMO and DSEC report future prediction mIoU/R@0.5, current-instant pIoU, runtime/prediction-age, ablations, and two downstream applications (VO with RAMP-VO and ViT token pruning).

Significance. If the results hold, the method is a practical, real-time event filtering mechanism with clear downstream utility. Strengths include a multi-task architecture with ATC, causal future prediction, a high inference rate, evaluation on two datasets, and downstream applications. However, headline claims need correction: the 13% ATE is an upper bound, the 67% improvement is against a custom retrained baseline on a specific metric, and the formal/experimental support for 'anticipatory suppression before they occur' is incomplete. These issues are fixable with rewording and additional event-level evaluation.

major comments (4)

[Abstract, Conclusion, and §IV-G] The abstract and conclusion claim a 13% ATE reduction, but §IV-G reports a mean ATE of 0.21 m (−8.7%) and explicitly describes −13% as the maximum achievable gain demonstrated by RAMP-VO† with ground-truth masks. The 13% figure is an upper bound, not the method's result. This misreporting appears in the abstract, introduction, and conclusion and should be corrected.
[§III-A, Eq. (3*), Fig. 2] The formal operator Sθ in Eq. (3*) selects events from the past window E[t−Δt,t) after warping the mask to time t, while Fig. 2 and §III-B describe suppression of future events E[t,t+Δtp). No formal definition or experiment measures event-level labeling in the future window; only mask mIoU at the horizon endpoint is reported. The central claim of anticipatory suppression 'before they occur' is therefore not directly supported by the paper's mathematics or metrics.
[§IV-B, Table I, Abstract] The advertised 67% segmentation improvement is the relative gain in R@0.5 against the self-retrained EVIMO† causal variant. Against the non-causal EVIMO* it is ~45%, and against current-instant point IoU in Table II the average gain over EVIMO is ~9%. The abstract's unqualified 'outperforming previous state-of-the-art methods by 67%' should specify the metric and baseline to avoid overstatement.
[§III-A, Eq. (4), §IV-F] The method assumes a single time-independent mask warped by one flow field is sufficient for the whole suppression window. The paper's own horizon ablation (Fig. 6) shows monotonic mIoU decay with Δtp, steepest on fast sequences, but this is endpoint mask accuracy, not suppression accuracy inside the window. Since events occur at intermediate times, a single warp cannot represent evolving IMO support under occlusions or non-smooth motion. Please provide event-level suppression evaluation within [t, t+Δtp) or explicitly state this as a limitation.

minor comments (4)

[Abstract and Table III] The abstract claims 'less than 1 GB of memory usage', but no memory measurement is reported in Table III or elsewhere. Add memory profiling or remove the claim.
[References and Table II] Citation inconsistencies: §IV-C cites EVIMO as [12] while Table II uses [10]; GConv is cited as [7] in Related Work and [13] in the experiments; Table I labels EV-IMO variants as [30] rather than [10]. Please unify references.
[§III-A and §III-D] Typos: 'misalingment' in §III-A and 'inspiried' in §III-D should be corrected.
[§IV-H] The 83% ViT acceleration is derived from Fig. 7 (approximately 12 to 22 images/s). The text says 'approximately 10 FPS' increase; please state the relative percentage explicitly and note that the masks are assumed to be precomputed at no extra time cost.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: future-mask predictions are genuine supervised forecasts evaluated on held-out data, and the architecture is not derived from its own claims.

full rationale

The paper's central pipeline is an end-to-end network that jointly predicts a current IMO mask and a future dense flow field, then warps the mask forward. The future masks are directly supervised by future ground-truth masks (Eq. 9), and the flow is trained with supervised losses where available (DSEC static regions) plus an unsupervised contrast-maximization objective. Evaluations on EVIMO/DSEC are performed on held-out test data against independent baselines, and the metrics (mIoU, R@0.5, pIoU, ATE) are computed from ground-truth annotations. The time-independence assumption in Section III-A (Eqs. 3-4) is a stated tractability approximation rather than an equivalence that makes the output equal to the input; the paper explicitly discusses the resulting horizon-dependent degradation in Fig. 6, which is consistent with genuine extrapolation difficulty, not with a fitted quantity being renamed a prediction. Self-citations to RAMP-VO, SViT, DSEC, and contrast-maximization literature are used as baselines or borrowed methodological components, but they are not invoked as uniqueness theorems or as the justification for the central claim. No equation in the paper reduces by construction to its own training targets, and no fitted parameter is presented as an independent prediction. Therefore no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No invented physical entities. The central contribution is a learned architecture; the free parameters are hand-set loss weights and input discretization choices rather than physical constants. The key load-bearing assumption is the single-mask-plus-single-flow warping model over the event window.

free parameters (3)

Loss weights (λ_future_mask=1.0, λ_flow=0.1, λ_smooth=0.05, w_sup=1.0, w_unsup=1.0) = See Appendix A-D
Chosen by hand; the paper states robustness to variations, but they are still free hyperparameters in the loss that the reported performance depends on.
Event aggregation window and voxel bins = 50 ms, 2 bins
Input discretization chosen without a systematic search; the appendix's event-count sensitivity experiment (Fig. 9) shows the method's accuracy depends on event density, so this choice affects the claims.
Prediction horizon randomization = Δt_p sampled uniformly in [0,100] ms
Training distribution over forecast horizons is a design choice that determines the reported 100 ms anticipatory capability and the prediction-age figures.

assumptions (4)

domain assumption Events in a scene can be partitioned into exactly two classes — ego-motion background and independently moving objects — with no other significant cause of events.
Section I framing; the entire suppression task and Eq. (2) assume this dichotomy.
ad hoc to paper A single time-independent binary mask over [t−Δt,t), warped by one predicted dense flow field, is a sufficient spatio-temporal model for suppression (Eqs. 3–4).
Explicitly assumed in Section III-A; false for rapidly deforming or occluded IMOs within the window.
standard math Contrast maximization on the Image of Warped Events is a valid training signal for the flow head when ground-truth flow is absent or invalid.
Taken from prior work [17],[25],[29]; used in L_unsup on EVIMO and on DSEC IMO regions where flow GT is not reliable.
domain assumption A two-bin voxel grid over 50 ms preserves enough temporal structure for 100 ms forecasting.
Used in Section IV-A without comparison to finer temporal discretizations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Motion-aware Event Suppression for Event Cameras." pith.science (2026). https://pith.science/paper/MRXIWQJ4

@misc{pith2026260223204,
  author       = {Pith},
  title        = {Pith review of: Motion-aware Event Suppression for Event Cameras},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRXIWQJ4}},
  note         = {Machine review of arXiv:2602.23204}
}

read the original abstract

Event cameras report asynchronously per-pixel brightness changes with microsecond latency, encoding dynamic visual information as a sparse stream of events. However, their extreme temporal resolution floods perception systems with entangled events from ego-motion and independently moving objects (IMOs), which existing solutions fail to efficiently decouple, relying instead on prohibitive dense 3D reconstructions or limited hand-tuned filters. In this work, we introduce the first framework for Motion-aware Event Suppression, which learns to filter events triggered by IMOs and ego-motion in real time. Our model jointly segments IMOs in the current event stream while predicting their future motion, enabling anticipatory suppression of dynamic events before they occur. Our lightweight architecture achieves 173 Hz inference on consumer-grade GPUs with less than 1 GB of memory usage, outperforming previous state-of-the-art methods on the challenging EVIMO benchmark by 67\% in segmentation accuracy while operating at a 53\% higher inference rate. Moreover, we demonstrate significant benefits for downstream applications: our method accelerates Vision Transformer inference by 83\% via token pruning and improves event-based visual odometry accuracy, reducing Absolute Trajectory Error (ATE) by 13\%.

Figures

Figures reproduced from arXiv: 2602.23204 by the authors.

**Figure 1.** Our method disentangles ego-motion events triggered by independently moving objects (IMOs). (a) We jointly learn to segment IMOs and predict [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Overview of the Anticipatory Motion Suppression pipeline A stack of input events E[t−∆t,t) is processed by a network featuring a series of recurrent blocks Ee followed by our proposed attention-based time conditioning (ATC) and task-specific decoders. The network jointly predicts a binary dynamic-object mask Mt and a future dense optical flow ψt→t+∆tp . The predicted flow is used in a flow warping module (FW) to for… view at source ↗

**Figure 3.** Overview of Attention-based Time Conditioning (ATC). Target time ∆tp is mapped via Positional Encoding (PE) to an embedding of size C, matching the spatial feature channels. These temporal features are broadcasted to form a Query, while flattened spatial features serve as the Key and Value. Cross-attention modulates the spatial features based on the temporal query, yielding the time-conditioned embedding E′ . To lev… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Anticipatory motion suppression on EVIMO (a) and DSEC (b). Top [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Dynamic object masks for SViT token pruning. Left to right: SViT-0 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Mean IoU (mIoU) versus prediction horizon [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 8.** Figure 8: Visualization of anticipatory motion suppression on EVIMO (a-b) and DSEC (c-d). Top row: overlaid events accumulated over [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

**Figure 9.** Figure 9: Segmentation accuracy (mIoU) as a function of the number of input [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 6 linked inside Pith

[1]

Event-based kilohertz eye tracking using coded differential lighting,

T. Stoffregen, H. Daraei, C. Robinson, and A. Fix, “Event-based kilohertz eye tracking using coded differential lighting,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 2515–2523

2022
[2]

Low-latency automotive vision with event cameras,

D. Gehrig and D. Scaramuzza, “Low-latency automotive vision with event cameras,”Nature, vol. 629, no. 8014, pp. 1034–1040, 2024

2024
[3]

Dsec: A stereo event camera dataset for driving scenarios,

M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “Dsec: A stereo event camera dataset for driving scenarios,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4947–4954, 2021

2021
[4]

Out of the room: Generalizing event-based dynamic motion segmen- tation for complex scenes,

S. Georgoulis, W. Ren, A. Bochicchio, D. Eckert, Y . Li, and A. Gawel, “Out of the room: Generalizing event-based dynamic motion segmen- tation for complex scenes,” in2024 International Conference on 3D Vision (3DV). IEEE, 2024, pp. 442–452

2024
[5]

Un-evmoseg: Unsuper- vised event-based independent motion segmentation,

Z. Wang, J. Guo, and K. Daniilidis, “Un-evmoseg: Unsuper- vised event-based independent motion segmentation,”arXiv preprint arXiv:2312.00114, 2023

arXiv 2023
[6]

A multi-scale recurrent framework for motion segmentation with event camera,

S. Zhang, L. Sun, and K. Wang, “A multi-scale recurrent framework for motion segmentation with event camera,”IEEE Access, vol. 11, pp. 80 105–80 114, 2023

2023
[7]

Learning visual motion segmentation using event surfaces,

A. Mitrokhin, Z. Hua, C. Fermuller, and Y . Aloimonos, “Learning visual motion segmentation using event surfaces,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 14 414–14 423

2020
[8]

Spikems: Deep spiking neural network for motion segmentation,

C. M. Parameshwara, S. Li, C. Ferm ¨uller, N. J. Sanket, M. S. Evanusa, and Y . Aloimonos, “Spikems: Deep spiking neural network for motion segmentation,” in2021 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS). IEEE, 2021, pp. 3414–3420

2021

Show all 35 references

[9]

Retina-inspired object motion segmentation for event-cameras,

V . Clerico, S. Snyder, A. Lohia, M. Abdullah-Al Kaiser, G. Schwartz, A. Jaiswal, and M. Parsa, “Retina-inspired object motion segmentation for event-cameras,”arXiv preprint arXiv:2408.09454

arXiv
[10]

Ev-imo: Motion segmentation dataset and learning pipeline for event cameras,

A. Mitrokhin, C. Ye, C. Ferm ¨uller, Y . Aloimonos, and T. Delbruck, “Ev-imo: Motion segmentation dataset and learning pipeline for event cameras,” in2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019, pp. 6105–6112

2019
[11]

Evdodgenet: Deep dynamic obstacle dodging with event cameras,

N. J. Sanket, C. M. Parameshwara, C. D. Singh, A. V . Kuruttukulam, C. Ferm ¨uller, D. Scaramuzza, and Y . Aloimonos, “Evdodgenet: Deep dynamic obstacle dodging with event cameras,” in2020 IEEE Interna- tional Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 10 651–10 657

2020
[12]

Event-based moving object detection and tracking,

A. Mitrokhin, C. Ferm ¨uller, C. Parameshwara, and Y . Aloimonos, “Event-based moving object detection and tracking,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1–9

2018
[13]

Event-based motion segmentation with spatio-temporal graph cuts,

Y . Zhou, G. Gallego, X. Lu, S. Liu, and S. Shen, “Event-based motion segmentation with spatio-temporal graph cuts,”IEEE transactions on neural networks and learning systems, vol. 34, no. 8, pp. 4868–4880, 2021

2021
[14]

Dynamic obstacle avoid- ance for quadrotors with event cameras,

D. Falanga, K. Kleber, and D. Scaramuzza, “Dynamic obstacle avoid- ance for quadrotors with event cameras,”Science Robotics, vol. 5, no. 40, p. eaaz9712, 2020

2020
[15]

0-mms: Zero-shot multi-motion segmentation with a monocular event camera,

C. M. Parameshwara, N. J. Sanket, C. D. Singh, C. Ferm ¨uller, and Y . Aloimonos, “0-mms: Zero-shot multi-motion segmentation with a monocular event camera,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9594–9600

2021
[16]

Event-based motion segmentation by motion compensation,

T. Stoffregen, G. Gallego, T. Drummond, L. Kleeman, and D. Scara- muzza, “Event-based motion segmentation by motion compensation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 7244–7253

2019
[17]

A unifying contrast maximization framework for event cameras, with applications to mo- tion, depth, and optical flow estimation,

G. Gallego, H. Rebecq, and D. Scaramuzza, “A unifying contrast maximization framework for event cameras, with applications to mo- tion, depth, and optical flow estimation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

2018
[18]

Wandering around: A bioinspired approach to visual attention through object motion sensitivity,

G. D’Angelo, V . Clerico, C. Bartolozzi, M. Hoffmann, P. M. Furlong, and A. Hadjiivanov, “Wandering around: A bioinspired approach to visual attention through object motion sensitivity,”Neuromorphic Com- puting and Engineering, vol. 5, no. 2, p. 024019, 2025

2025
[19]

Ev-flownet: Self- supervised optical flow estimation for event-based cameras,

A. Z. Zhu, L. Yuan, K. Chaney, and K. Daniilidis, “Ev-flownet: Self- supervised optical flow estimation for event-based cameras,”arXiv preprint arXiv:1802.06898, 2018

2018 arXiv
[20]

Unsupervised learning of dense optical flow, depth and egomotion from sparse event data,

C. Ye, A. Mitrokhin, C. Ferm ¨uller, J. A. Yorke, and Y . Aloimonos, “Unsupervised learning of dense optical flow, depth and egomotion from sparse event data,”arXiv preprint arXiv:1809.08625, 2018

2018 arXiv
[21]

E-raft: Dense optical flow from event cameras,

M. Gehrig, M. Millh ¨ausler, D. Gehrig, and D. Scaramuzza, “E-raft: Dense optical flow from event cameras,” in2021 International Con- ference on 3D Vision (3DV). IEEE, 2021, pp. 197–206

2021
[22]

Taming contrast maximization for learning sequential, low-latency, event-based optical flow,

F. Paredes-Vall ´es, K. Y . Scheper, C. De Wagter, and G. C. De Croon, “Taming contrast maximization for learning sequential, low-latency, event-based optical flow,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 9695–9705

2023
[23]

Secrets of event-based optical flow,

S. Shiba, Y . Aoki, and G. Gallego, “Secrets of event-based optical flow,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 628– 645

2022
[24]

Learning to reconstruct hdr images from events, with applications to depth and flow prediction,

M. Mostafavi, L. Wang, and K.-J. Yoon, “Learning to reconstruct hdr images from events, with applications to depth and flow prediction,” International Journal of Computer Vision, vol. 129, no. 4, pp. 900–920, 2021

2021
[25]

Taming contrast maximization for learning sequential, low- latency, event-based optical flow,

F. Paredes-Vall ´es, K. Y . W. Scheper, C. De Wagter, and G. C. H. E. de Croon, “Taming contrast maximization for learning sequential, low- latency, event-based optical flow,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 9695–9705

2023
[26]

Delving deeper into con- volutional networks for learning video representations,

N. Ballas, L. Yao, C. Pal, and A. Courville, “Delving deeper into con- volutional networks for learning video representations,”arXiv preprint arXiv:1511.06432, 2015

2015 arXiv
[27]

De- terministic edge-preserving regularization in computed imaging,

P. Charbonnier, L. Blanc-F ´eraud, G. Aubert, and M. Barlaud, “De- terministic edge-preserving regularization in computed imaging,”IEEE Transactions on image processing, vol. 6, no. 2, pp. 298–311, 1997

1997
[28]

Dice loss for data-imbalanced nlp tasks,

X. Li, X. Sun, Y . Meng, J. Liang, F. Wu, and J. Li, “Dice loss for data-imbalanced nlp tasks,”arXiv preprint arXiv:1911.02855, 2019

1911 arXiv
[29]

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

G. Gallego, M. Gehrig, and D. Scaramuzza, “Focus is all you need: Loss functions for event-based vision,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

2019
[30]

Unsupervised learning of dense optical flow, depth and egomotion with event-based sensors,

C. Ye, A. Mitrokhin, C. Ferm ¨uller, J. A. Yorke, and Y . 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). IEEE, 2020, pp. 5831–5838

2020
[31]

Deep visual odometry with events and frames,

R. Pellerito, M. Cannici, D. Gehrig, J. Belhadj, O. Dubois-Matra, M. Casasco, and D. Scaramuzza, “Deep visual odometry with events and frames,” inIEEE/RSJ International Conference on Intelligent Robots (IROS), June 2024

2024
[32]

Revisiting token pruning for object detection and instance segmentation,

Y . Liu, M. Gehrig, N. Messikommer, M. Cannici, and D. Scara- muzza, “Revisiting token pruning for object detection and instance segmentation,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 2658–2668

2024
[33]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017. Appendix for: Motion-aware Event Suppression for Event Cameras Roberto Pellerito,...

2017
[34]

Supervised Loss Component (L sup):The supervised loss term is a combination of losses for the dynamic object seg- mentation and optical flow prediction tasks, applied wherever ground truth is available. a) Segmentation Loss (L mask):For the segmentation task, we use a combined...
[35]

The core idea is that a correct optical flow field, when used to warp events over a time interval, will ”deblur” the event stream by mapping moving points back to their origin

Unsupervised Flow Loss (L unsup):For the optical flow forecasting task, we incorporate a powerful, unsupervised loss based on the principle of contrast maximization [29] and inspired by [25]. The core idea is that a correct optical flow field, when used to warp events over a t...

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Event-based kilohertz eye tracking using coded differential lighting,

T. Stoffregen, H. Daraei, C. Robinson, and A. Fix, “Event-based kilohertz eye tracking using coded differential lighting,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 2515–2523

2022

[2] [2]

Low-latency automotive vision with event cameras,

D. Gehrig and D. Scaramuzza, “Low-latency automotive vision with event cameras,”Nature, vol. 629, no. 8014, pp. 1034–1040, 2024

2024

[3] [3]

Dsec: A stereo event camera dataset for driving scenarios,

M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “Dsec: A stereo event camera dataset for driving scenarios,”IEEE Robotics and Automation Letters, vol. 6, no. 3, pp. 4947–4954, 2021

2021

[4] [4]

Out of the room: Generalizing event-based dynamic motion segmen- tation for complex scenes,

S. Georgoulis, W. Ren, A. Bochicchio, D. Eckert, Y . Li, and A. Gawel, “Out of the room: Generalizing event-based dynamic motion segmen- tation for complex scenes,” in2024 International Conference on 3D Vision (3DV). IEEE, 2024, pp. 442–452

2024

[5] [5]

Un-evmoseg: Unsuper- vised event-based independent motion segmentation,

Z. Wang, J. Guo, and K. Daniilidis, “Un-evmoseg: Unsuper- vised event-based independent motion segmentation,”arXiv preprint arXiv:2312.00114, 2023

arXiv 2023

[6] [6]

A multi-scale recurrent framework for motion segmentation with event camera,

S. Zhang, L. Sun, and K. Wang, “A multi-scale recurrent framework for motion segmentation with event camera,”IEEE Access, vol. 11, pp. 80 105–80 114, 2023

2023

[7] [7]

Learning visual motion segmentation using event surfaces,

A. Mitrokhin, Z. Hua, C. Fermuller, and Y . Aloimonos, “Learning visual motion segmentation using event surfaces,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 14 414–14 423

2020

[8] [8]

Spikems: Deep spiking neural network for motion segmentation,

C. M. Parameshwara, S. Li, C. Ferm ¨uller, N. J. Sanket, M. S. Evanusa, and Y . Aloimonos, “Spikems: Deep spiking neural network for motion segmentation,” in2021 IEEE/RSJ International Conference on Intelli- gent Robots and Systems (IROS). IEEE, 2021, pp. 3414–3420

2021