Research map / Detection and segmentation / CNN features
EGSST: Event-based Graph Spatiotemporal Sensitive Transformer for Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EGSST, a graph-based event-camera object detector that builds connected subgraphs from event graphs, gates temporal attention with a spatiotemporal sensitivity module, and uses a multi-scale linear ViT backbone with RT-DETR or YOLOX heads.
Problem
Object detection from event camera data, where sparse asynchronous events are poorly suited to frame-based methods and temporal processing wastes computation on slow-moving content.
Approach
Model events as a spatiotemporal graph split into connected subgraphs, measure relative dynamics (EGM/ELM) with an SSM, and adaptively activate a temporal attention controller (TAC) only when dynamics warrant it, combined with a lightweight multi-scale linear ViT.
Key results
- EGSST-E reaches 49.6% mAP on Gen1 at 6.0 ms and 50.2% mAP on 1 Mpx at 6.3 ms with 12.3M parameters.
- EGSST-B reaches 44.6% mAP on Gen1 at 4.6 ms with only 3.5M parameters.
- In the ablation on Gen1, TAC adaptive gives 49.6% mAP at 6.0 ms versus 51.5% at 11.1 ms when TAC is always active, trading a small accuracy loss for lower latency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adaptively activate TAC using SSM instead of keeping TAC inactive (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 42.1 -> 49.6 | Helped |
| adaptively activate TAC using SSM instead of always activating TAC (ablation) | EGSST-E · Gen1 | time per batch (ms): 11.1 -> 6.0 | Helped |
| always activate TAC instead of adaptive TAC (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 49.6 -> 51.5 | Helped |
| always activate TAC instead of adaptive TAC (ablation) | EGSST-E · Gen1 | time per batch (ms): 6.0 -> 11.1 | Hurt |
| add CNN feature-extraction module (ablation) | EGSST-B · Gen1 | mAP@0.5:0.95: +0.4 | Helped |
| add CNN feature-extraction module (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: +0.5 | Helped |
| add CNN feature-extraction module (ablation) | EGSST-E · Gen1 | time per batch (ms): +0.1 | Hurt |
| add dynamic label augmentation (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 38.3 -> 46.6 | Helped |
| add horizontal-flip augmentation (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 38.3 -> 41.8 | Helped |
| add zoom-in augmentation (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 38.3 -> 43.3 | Helped |
| add zoom-out augmentation (ablation) | EGSST-E · Gen1 | mAP@0.5:0.95: 38.3 -> 42.1 | Helped |
| increase input event count (ablation) | EGSST-E (10T) · Gen1, 2,000 events | mAP: 33.1 -> 39.4 | Helped |
Surprising findings
- Integrating a simple CNN module into the otherwise event-based pipeline improved mAP by about 0.4-0.5 points at minimal cost, though the model is described as fully event-based.
Limitations
- GNN deployment is harder than ViT deployment, and the authors note further work is needed on GNN deployment efficiency.
- No error bars or statistical significance reported.
- Relies on a CNN layer for the best results, so the pipeline is not purely event-based in the reported configuration.
Experimental setup
Datasets: Gen1 Automotive Detection Dataset 1 Megapixel Automotive Detection Dataset
Benchmarks: Gen1 Automotive Detection Dataset 1 Megapixel Automotive Detection Dataset
Metrics: mAP@0.5:0.95 inference time per batch (ms) parameter count (M)
Baselines: RRC-Events AEGNN Spiking DenseNet ERGO-12 RED ASTMet AEC + DETR AEC + YOLOv5 RVT-B GET-T S4D-ViT-B S5-ViT-B
Builds on: Efficient ViT Graph Attention Network RT-DETR YOLOX AEGNN
Closest papers by meaning
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.88)
- Scene Adaptive Sparse Transformer for Event-based Object Detection (CVPR 2024 · similarity 0.87)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.86)
- Hybrid Spiking Vision Transformer for Object Detection with Event Cameras (ICML 2025 · similarity 0.85)
- SMamba: Sparse Mamba for Event-based Object Detection (AAAI 2025 · similarity 0.85)
- Motion-adaptive Transformer for Event-based Image Deblurring (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.