acceptodds
Withdrawn from ICLR 2027

CSEG: A COMPACT STATEFUL EVENT GRAPH REPRESENTATION FOR EFFICIENT VISUAL PROCESSING

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The sparse, asynchronous output of event cameras is naturally suited to graph representations, yet treating each event as a node makes such graphs computationally expensive. Discarding events reduces this cost but inevitably loses information. Instead, we introduce CSEG, a compact stateful event graph representation that significantly reduces graph size while maintaining competitive detection accuracy. CSEG constructs a compact graph by retaining only a subset of event nodes, selected according to polarity and temporal information. For each dropped event node, local message differences guide the selection of a retained node as its fusion target. Its temporal information is then fused into this node through a temporal state update. We evaluate CSEG on event-based object detection using the Gen1 and DSEC-Det datasets. Compared with existing event graph methods, CSEG achieves competitive detection accuracies while retaining only 42.56% and 44.23% of event nodes on Gen1 and DSEC-Det datasets, respectively. Compared with the event graph baseline, CSEG achieves a 2.6x acceleration in inference, demonstrating its computational efficiency.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
event camera, low latency, event graph

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7A8WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the computational efficiency of graph representations for event cameras. Existing event-graph methods typically represent each individual event as a graph node, which causes the graph size and computational overhead to scale with the number of events. We propose CSEG, a graph compression strategy that retains a subset of event nodes while fusing the temporal contributions of discarded events into the kept nodes. Specifically, CSEG determines whether to preserve an incoming event based on the temporal state change of existing nodes, and selects fusion targets using local message differences from the first graph convolution layer. The framework further leverages cached features and incremental updates for asynchronous graph inference. Evaluated on Gen1 and DSEC-Det, our method retains approximately 42.56% and 44.23% of event nodes respectively, and achieves an mAP of 38.7 and 42.17. Under the measurement setup with cached image features and 1 ms event batching, we report around 2.6x speedup compared with DAGr.

Critical strengths

S1. For the information loss incurred in graph compression, we propose a representation that is well-founded and amenable to incremental updates. CSEG accumulates the temporal contributions of discarded events into retained nodes instead of simply dropping them. This design targets the practical bottleneck of computational overhead in event graphs and exhibits good compatibility with asynchronous inference. On both datasets, competitive detection performance is maintained while retaining fewer than half of the event nodes, demonstrating the practical value of this research direction. S2. Ablation studies provide direct evidence for the effectiveness of temporal information fusion. As shown in Table 3, under the same node retention ratio of 44.23%, CSEG improves from 41.31 mAP without temporal fusion to 42.17 mAP for the full model. The random retention baseline also benefits when fusion is enabled.Temporal activity reconstruction results in Table 4 further verify that fusion preserves more aggregated temporal information. Collectively, these results demonstrate that temporal information fusion contributes meaningfully to the final performance.

Critical weaknesses

W1. The current comparisons do not sufficiently isolate the gains brought by the proposed method itself. W2. Ambiguities remain in the accounting criteria of efficiency experiments, which affect the core conclusion regarding speedup.

Questions for authors
  1. What is the performance gain of message-guided fusion over nearest-neighbor fusion and random fusion under strictly matched node budgets and training conditions? When varying the retention threshold, is the trade-off between accuracy and practical latency stable?
Reviewer_N5BdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CSEG, a compact stateful event graph representation for efficient event-based visual processing. The method reduces graph size through state-guided node retention and message-guided fusion target selection. Instead of directly discarding events, CSEG aggregates their temporal contributions into retained nodes. Lookup tables and cached graph features are further used to support incremental graph updates and asynchronous inference. The method is evaluated on event-based object detection using Gen1 and DSEC-Det. The results show that CSEG retains approximately 43–44% of event nodes while achieving competitive detection accuracy and a reported 2.6× inference speedup over DAGr.

Critical strengths
  1. Reducing the computational overhead of event graphs is an important problem, especially for high-frequency asynchronous event streams. CSEG provides a practical approach to reducing graph size while retaining aggregate temporal information. The reported improvements in inference latency and computational cost are encouraging.
Critical weaknesses

1.Limited novelty and insufficient comparison with related representations.

1.1. The proposed node retention, local merging, and feature aggregation are closely related to established graph pooling and coarsening methods, including DiffPool [1], SAGPool [2], EdgePool [3], and ASAP [4]. Although CSEG adapts these ideas to asynchronous event processing, its methodological novelty is not sufficiently distinguished from existing approaches.

1.2. Moreover, recent event representation methods are not adequately discussed or compared, including Verma et al. (WACV 2026) [5], Spike-driven Discrete Aggregation (CVPR 2026) [6], and the point-based EventMamba (TPAMI 2025) [8]. In particular, EventMamba demonstrates an alternative approach to processing sparse event clouds without explicit graph construction.

1.3. The advantages of CSEG over such point-based representations remain unclear. The current ablations focus mainly on internal components, rather than establishing whether the proposed compression and state fusion mechanisms outperform simpler event aggregation or graph coarsening strategies.

  1. Insufficient experimental validation, scalability analysis, and reproducibility.

2.1. The evaluation is restricted to object detection on Gen1 and DSEC-Det, which is insufficient to establish CSEG as a general representation for visual processing. Its effectiveness on classification, semantic segmentation, or optical flow remains unexplored.

2.2. An important missing benchmark is the 1Mpx Automotive Detection dataset [9]. Its 1280×720 event resolution would provide a more convincing evaluation of scalability, memory efficiency, and information preservation under high-resolution inputs.

2.3. The detection results are also mixed. On Gen1, CSEG achieves 38.7 mAP, below SAST (48.2) and RVT (47.2). On DSEC-Det, it only marginally improves over DAGr (42.17 vs. 41.9 mAP). More comprehensive comparisons with recent detection methods, including FlexEvent [7], would strengthen the evaluation.

2.4. The efficiency analysis is also incomplete. The reported latency is measured on only 24 DSEC test windows with cached image features. Moreover, Appendix C.2 describes extensive CUDA optimizations and feature caching, but their contributions are not separated from those of graph compression. Training efficiency, convergence behavior, and statistical variance are not sufficiently reported. Finally, only a runnable demo is available for review, limiting independent reproducibility. Overall, the reported speedup is useful, but the current evidence does not sufficiently establish the novelty, generality, or scalability of the proposed representation.

References
[1] Ying et al. Hierarchical Graph Representation Learning with Differentiable Pooling. NeurIPS, 2018.
[2] Lee et al. Self-Attention Graph Pooling. ICML, 2019.
[3] Diehl. Edge Contraction Pooling for Graph Neural Networks. 2019.
[4] Ranjan et al. ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations. AAAI, 2020.
[5] Verma et al. Event-based Graph Representation with Spatial and Motion Vectors for Asynchronous Object Detection. WACV, 2026.
[6] Li et al. Spike-driven Discrete Aggregation for Event-based Object Detection. CVPR, 2026.
[7] Lu et al. FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies. NeurIPS, 2025.
[8] Ren et al. Rethinking Efficient and Effective Point-Based Networks for Event Camera Classification and Regression. IEEE TPAMI, 2025.
[9] Perot et al. Learning to Detect Objects with a 1 Megapixel Event Camera. NeurIPS, 2020. \

Questions for authors

Q1 (Novelty): How does CSEG fundamentally differ from existing graph pooling and coarsening methods [1–4]? Could the authors compare against simpler local node merging or aggregation strategies?

Q2 (Recent methods): Could the authors discuss and compare CSEG with recent event graph and aggregation methods, particularly Verma et al. [5] and Spike-driven Discrete Aggregation [6]? How does CSEG compare with EventMamba [8] or similar point-based event representations? What advantages justify explicit graph construction and stateful node fusion in terms of accuracy, temporal information preservation, and computational efficiency?

Q3 (Scope): The title suggests a general event representation for visual processing, but experiments only cover object detection. Could the authors justify this claim or narrow the title accordingly?

Q4 (Generalization): Could the authors provide results on at least one additional task, such as event classification or optical flow? This would help assess whether the compressed representation preserves information beyond detection.

Q5 (High-resolution scalability): Why is the 1Mpx dataset [9] not included? Could the authors evaluate detection accuracy, node retention, latency, and memory consumption at 1280×720 resolution?

Q6 (Fair comparisons): Could the authors compare different event representations under a common detection backbone and comparable computational budgets? This would better isolate the contribution of CSEG from the detector architecture.

Q7 (Robustness and trade-offs): How does performance change with event density, noise, camera motion, and retention ratio? Could the authors report accuracy, latency, FLOPs, and memory across different compression settings?

Q8 (End-to-end efficiency): Could the authors report end-to-end latency and sustained throughput, including preprocessing, graph construction, temporal fusion, image feature extraction, and detection? The current measurements use cached image features, and the reported 5.20 ms processing time for 1 ms event batches raises questions about real-time operation. How much of the reported 2.6× speedup comes from graph compression versus CUDA optimizations, feature caching, and graph replay? A breakdown under matched implementation conditions would clarify the algorithmic contribution.

Q9 (Training and optimization Details): Could the authors provide the exact training loss formulation, training/validation loss curves, training time and memory, and an analysis of the non-differentiable retention and fusion decisions? How sensitive is training to the LUT update frequency? Could the authors provide complete training configurations, random seeds, multi-run statistics, and the full training and asynchronous inference code? Since the improvement over DAGr on DSEC-Det is only 0.27 mAP, statistical reliability is particularly important. It would also be useful to verify the consistency between synchronous training and asynchronous inference.

References
[1] Ying et al. Hierarchical Graph Representation Learning with Differentiable Pooling. NeurIPS, 2018.
[2] Lee et al. Self-Attention Graph Pooling. ICML, 2019.
[3] Diehl. Edge Contraction Pooling for Graph Neural Networks. 2019.
[4] Ranjan et al. ASAP: Adaptive Structure Aware Pooling for Learning Hierarchical Graph Representations. AAAI, 2020.
[5] Verma et al. Event-based Graph Representation with Spatial and Motion Vectors for Asynchronous Object Detection. WACV, 2026.
[6] Li et al. Spike-driven Discrete Aggregation for Event-based Object Detection. CVPR, 2026.
[7] Lu et al. FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies. NeurIPS, 2025.
[8] Ren et al. Rethinking Efficient and Effective Point-Based Networks for Event Camera Classification and Regression. IEEE TPAMI, 2025.
[9] Perot et al. Learning to Detect Objects with a 1 Megapixel Event Camera. NeurIPS, 2020. \

Reviewer_62X1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes CSEG, a compact stateful event graph representation that resolves the trade-off between graph scale and feature integrity. CSEG downsizes graphs by selectively retaining event nodes based on polarity and temporal cues. For discarded events, it fuses their temporal information into suitable reserved nodes according to local message differences through temporal state updates, effectively preserving critical information despite node reduction. Evaluated on the Gen1 and DSEC-Det event-based object detection datasets, CSEG retains only 42.56% and 44.23% of original event nodes respectively while achieving competitive detection accuracy. It also delivers a 2.6× inference speedup over the baseline event graph method, demonstrating its high computational efficiency for event vision tasks.

Critical strengths

1.The idea seems reasonable and interesting for the event stream based representation learning. 2.The experimental results on two object detection benchmark datasets fully validated the effectiveness of the proposed model on the efficiency improvement. 3.The writing of this paper is good and easy to following all through this paper.

Critical weaknesses

1.The related works mentioned in section 2 seems relatively old, these works are ranging from year 2017 to 2024. Only one paper published in year 2025, and 0 in year 2026. I think the review on the related works is not up to date. 2.The experiments are conducted only on object detection tasks, which is not comprehensive for a new event representation. In my view, the event-based pattern recognition, visual object tracking, sign language translation are all suggested for a more in-depth study. 3.The source code will be released or not is not clearly described in this paper. 4.The compared models in Table 1 are relative old, no works published in year 2025 and year 2026. 5.The comparison between the newly proposed one is not compared with other event representations, such as event tensor, event frame, event voxel, event graph, event hyper-graph, etc.

Questions for authors

In addition to the weakness part, I think the authors should validate their method on more event-based datasets, and compare it with more event-based representations and tasks. Also, more related work published in the past three years should be highlighted.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.