Motion-aware Event Suppression for Event Cameras

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

cs.CVFeb 26, 2026

Motion-aware Event Suppression for Event Cameras

Authors: Roberto Pellerito, Nico Messikommer, Giovanni Cioffi, Marco Cannici, Davide Scaramuzza

Organizations: Robotics and Perception Group, University of Zurich, Switzerland

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 & tables

Fig. 1: Our method disentangles ego-motion events from events triggered by independently moving objects (IMOs). (a) We jointly learn to segment IMOs and predict dense optical flow for the next $\Delta t$ . Warping the mask forward yields an anticipated future mask, enabling suppression of future events. (b) Compared to baselines like EVIMO (baseline), our approach produces tighter masks with higher IoU and fewer false positives. (c) Event Suppression can be used for downstream tasks: (top) accelerating segmentation via motion-guided token pruning, and (bottom) improving visual odometry by filtering out dynamic IMO edges.

Fig. 2: Overview of the Anticipatory Motion Suppression pipeline A stack of input events $\mathcal{E}$ is processed by a network featuring a series of recurrent blocks $E_{e}$ followed by our proposed attention-based time conditioning (ATC). Raw features from events are fed to the transposed convolutional decoders $D_{M}$ for IMO mask prediction . Time conditioned features are used by a second transposed convolutional decoder $D_{\psi}$ for flow forecast prediction . The network finally predicts a binary dynamic-object mask $M_{t}$ and a future dense optical flow $\psi_{t\rightarrow t+\Delta t_{p}}$ . The predicted flow is used in a flow warping module (FW) to forecast the motion of dynamic regions. By propagating the mask forward in time, the system anticipates and suppresses future events $\mathcal{E}_{[t,t+\Delta t_{p})}$ corresponding to either independently moving objects or the static background, yielding a simplified event stream focused on just one specific type of motion.

Fig. 3: Overview of Attention-based Time Conditioning (ATC). Target time

\Delta t_{p}

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^{\prime}

Fig. 4: Anticipatory motion suppression on EVIMO (a) and DSEC (b). Top row: Events accumulated over $\Delta t_{p}=100$ ms with ground-truth IMO masks. Bottom row: Predicted separation of ego-motion (white) and IMO (green) events for the future 100 ms window based on the previous 50 ms of data. Our pipeline demonstrates: (a) robustness under extreme ego-motion with complex IMO shapes and (b) accurate motion anticipation for both distant and nearby vehicles in driving scenes.

	OMS [ 9 ]		EV-IMO † [ 10 ]		EV-IMO ∗ [ 10 ]		Ours
Sequence	mIoU	[email protected]	mIoU	[email protected]	mIoU	[email protected]	mIoU	[email protected]
Boxes	49.15	2.21	69.74	53.73	73.53	61.05	76.24	75.38
Floor	48.23	0	77.06	55.95	75.47	66.93	80.10	95.51
Wall	48.63	0.27	75.45	40.02	75.78	68.06	79.63	88.09
Table	48.73	0.79	78.41	48.37	75.26	61.23	79.71	90.95
Fast	45.97	0.69	63.96	33.03	64.66	38.94	68.07	67.26

TABLE I: Mean Intersection Over Union % (mIoU) and ratio of correctly segmented objects % ([email protected]) for future prediction on EVIMO dataset.

Algorithm	Input	Boxes	Floor	Wall	Table	Fast
Baseline CNN [ 5 ]	E	50	74	60	66	52
EV-IMO [ 10 ]	E	70	59	78	79	67
EVDodgeNet [ 11 ]	E	67	61	72	70	60
SpikeMS [ 8 ]	E	65	53	63	50	38
GConv [ 7 ]	E	60	55	80	51	39
EMSGC [ 13 ] Top 30%	E	24	18	24	55	43

TABLE II: Point-based Intersection over Union % (pIoU) for mask prediction at the current instant on EVIMO dataset. E indicates methods using events while I indicates the method using RGB frames.

Table 7

Fig. 5: Dynamic object masks for SViT token pruning. Left to right: SViT-0 and SViT-16 show increasing mask dilation. Larger dilation factors cover more area, including more background tokens and increasing system latency. This mask expansion introduces redundancy that significantly reduces inference frequency while capturing the same dynamic content of the scene.

FW	ATC	Boxes	Floor	Wall	Table	Fast
$\times$	$\times$	72.25	73.08	68.62	78.71	65.06
Linear	$\times$	73.88	75.41	71.01	78.02	61.77
$\checkmark$	$\times$	74.03	78.02	75.14	77.70	67.49
$\checkmark$	$\checkmark$	76.24	80.10	79.63	79.71	68.07

TABLE V: Mean Intersection Over Union % (mIoU) Comparison of Core Components of Our Framework on EVIMO dataset

Model	Boxes	Floor	Wall	Table	Fast
XS	47.12	47.82	46.26	47.57	50.74
S	68.72	80.07	75.96	73.93	58.52
M	72.68	80.00	75.04	79.02	64.59
L	76.24	80.10	79.63	79.71	68.07

TABLE VI: Mean Intersection Over Union % (mIoU) for Our Model at Different Scales On EVIMO dataset

Fig. 6: Mean IoU (mIoU) versus prediction horizon

\Delta t_{p}

(0–100 ms) on the EVIMO dataset for six sequences (box, fast, floor, table, tabletop, wall) and the macro-average total. Performance declines with increasing horizon, indicating that longer look-ahead yields less accurate segmentation. Faster scenes such as wall and fast degrade more rapidly than slower ones like box and tabletop, due to stronger non-smooth motion and occlusions that increase the effective prediction age.

Algorithm	Boxes	Floor	Wall	Table	Fast	Tabletop
RAMP-VO [ 33 ]	0.33	0.17	0.34	0.15	0.25	0.14
RAMP-VO ∗ [ 33 ]	0.29	0.15	0.33	0.15	0.22	0.12
RAMP-VO † [ 33 ]	0.28	0.15	0.30	0.15	0.22	0.11

TABLE VII: ATE [m] (Absolute Trajectory Error) Comparison of RAMP-VO on EVIMO Dataset

Fig. 7: Trade-off between segmentation accuracy (Seg AP) and inference frequency for different pruning ratios of the SViT model. As the pruning ratio increases (e.g., from SViT-full to SViT-0), the inference frequency improves, but the segmentation performance degrades.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Table 14

Fig. 8: Visualization of anticipatory motion suppression on EVIMO (a-b) and DSEC (c-d). Top row: overlaid events accumulated over $\Delta t_{p}=100\,\text{ms}$ with ground-truth masks for independently moving objects (IMO). Bottom row: events attributed to ego-motion (white) and to IMO (green), obtained by applying a motion mask to future $\Delta t_{p}=100\,\text{ms}$ events, with motion predicted from the previous $\Delta t=50\,\text{ms}$ events using our Anticipatory Motion Suppression pipeline. Panels: (a) Robustness under fast ego-motion and non-convex masks, recovering IMO events from a large stream of ego-motion events. (b) Performance in indoor, multi-object scenes, with arbitrary shaped non-convex objects. (c-d) Autonomous driving results, anticipating the motion of small moving objects, both far-away vehicles and nearby objects.

Panel of Figure 16 — Fig. 9: Qualitative results on DSEC under challenging lighting conditions. (a, b) In night scenes, our method isolates dynamic objects despite sparse events and sensor noise, though flickering headlights can slightly inflate pedestrian boundaries. (c, d) In low-contrast cloudy scenarios, vehicles are reliably segmented across various distances. However, objects generating minimal temporal contrast (e.g., the missing pedestrian on the far left) may lack sufficient motion cues for detection. Yellow masks indicate ground truth; green indicates our predictions.

Panel of Figure 17 — Fig. 10: (a) EVIMO events. (b) Our IMO prediction (green) vs. ground truth (purple). (c) Overexposed frame. (d) FlowSAM prediction from (c), green: prediction, purple: ground truth.

Fig. 11: Segmentation accuracy (mIoU) as a function of the number of input events

N

used to form the network input. The x-axis is logarithmic. At test time, instead of fixing a temporal window

\Delta t

, we fix an event budget

N

and measure performance. Accuracy is poor for very sparse inputs (

N<10^{3}

), improves steadily, and reaches its maximum around

N\approx 6\times 10^{4}

. Beyond this point, performance gradually declines as overly long temporal integration blurs object boundaries. The steady region (

10^{4}\!-\!10^{5}

events) indicates a broad operating plateau in which the method is both accurate and stable across different scene dynamics.

Alg.	$\leq$ 90	90-120	$\geq$ 120
EV-IMO [10]	77.8	76.9	66.9
Ours	82.3	79.7	72.2

Appendix

TABLE X: EVIMO (mIoU%) by Speed (px/s)

Explore similar work

Real-time Motion Segmentation with Event-based Normal Flow
Feb 24, 2026Sheng Zhong, Zhongyang Ren, Xiya Zhu +3Event-Based Vision
Neuromorphic Automation and Intelligence Lab (NAIL) at School of Artificial Intelligence and Robotics, Hunan University, Changsha, China. · Capital One. · University of Maryland, College Park.
E-TIDE: Fast, Structure-Preserving Motion Forecasting from Event Sequences
Mar 29, 2026Biswadeep Sen, Benoit R. Cottereau, Nicolas Cuperlier +1Event-Based Vision Efficient Neural Network Inference
National University of Singapore, Singapore · CerCo, CNRS UMR 5549, Universit´e de Toulouse, France · ETIS UMR8051, CY Cergy Paris Universit´e, ENSEA, CNRS, Cergy, France +1
PointEvent: Rethinking Event-based Tiny Object Detection via Serialized Motion Evidence Accumulation
Sep 17, 2026Zongze Wu, Baofeng Jia, Weiqi Yan +4Small Object Detection Efficient Object Detection
State key Lab of Extreme Environment Optoelectronic Dynamic Testing Technology and Instrument, Nanjing University of Science and Technology, China · Jiangsu Key Lab of Visual Sensing and Intelligent Perception, Nanjing University of Science and Technology, China · Fujian Key Laboratory of Urban Intelligent Sensing and Computing, Xiamen University, China

Feb 24, 2026J/K move · Enter open · S save

Real-time Motion Segmentation with Event-based Normal Flow

Sheng Zhong, Zhongyang Ren, Xiya Zhu, Dehao Yuan+2

Neuromorphic Automation and Intelligence Lab (NAIL) at School of Artificial Intelligence and Robotics, Hunan University, Changsha, China. · Capital One. · University of Maryland, College Park.

Event-Based Vision

Open paper AskPDF arXiv

Event-based cameras are bio-inspired sensors with pixels that independently and asynchronously respond to brightness changes at microsecond resolution, offering the potential to handle visual tasks in challenging scenarios. However, due to the sparse information content in individual events, directly processing the raw event data to solve vision tasks is highly inefficient, which severely limits the applicability of state-of-the-art methods in real-time tasks, such as motion segmentation, a fundamental task for dynamic scene understanding. Incorporating normal flow as an intermediate representation to compress motion information from event clusters within a localized region provides a more effective solution. In this work, we propose a normal flow-based motion segmentation framework for event-based vision. Leveraging the dense normal flow directly learned from event neighborhoods as input, we formulate the motion segmentation task as an energy minimization problem solved via graph cuts, and optimize it iteratively with normal flow clustering and motion model fitting. By using a normal flow-based motion model initialization and fitting method, the proposed system is able to efficiently estimate the motion models of independently moving objects with only a limited number of candidate models, which significantly reduces the computational complexity and ensures real-time performance, achieving nearly a 800x speedup in comparison to the open-source state-of-the-art method. Extensive evaluations on multiple public datasets fully demonstrate the accuracy and efficiency of our framework.

Abstract

Explore similar work

Real-time Motion Segmentation with Event-based Normal Flow

E-TIDE: Fast, Structure-Preserving Motion Forecasting from Event Sequences

PointEvent: Rethinking Event-based Tiny Object Detection via Serialized Motion Evidence Accumulation