Research map / Detection and segmentation / CNN features
SMamba: Sparse Mamba for Event-based 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 SMamba, a Mamba-based event object detector that adaptively discards non-event/noisy tokens via spatiotemporal continuity assessment and adds information-prioritized scanning and global channel interaction, reaching higher mAP with fewer FLOPs on Gen1, 1Mpx and eTram.
Problem
Transformer-based event-based object detectors process sparse, noisy non-event regions uniformly, incurring high computation; window-based sparsification sacrifices global modeling.
Approach
Scores tokens by spatiotemporal continuity of events (accumulated timestamps, Gaussian-smoothed neighborhood) to prune uninformative tokens, reorders tokens by information so high-information tokens are scanned close together (IPL-Scan), and applies S6 across channels using global spatial content (GCI).
Key results
- On Gen1, SMamba reaches 50.4 mAP with 2.4G backbone FLOPs and 16.1M params, matching ERGO-12's mAP at about 5% of its FLOPs.
- On 1Mpx, SMamba achieves 49.3 mAP vs 48.7 for SAST-CB at 7.4G FLOPs; on eTram it reaches 32.6 mAP vs 30.0 for SAST-CB.
- Sparsification cuts FLOPs by 23%, 22% and 31% over the baseline on Gen1, 1Mpx and eTram, with mAP gains of 0.4, 0.5 and 0.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the STCA spatiotemporal-continuity scoring method instead of variance, entropy, or a learnable scoring module (ablation) | SMamba · eTram | mAP: 30.8, 30.4, 31.2 -> 32.6 | Helped |
| combine Bidi-Scan with IPL-Scan (ablation) | SMamba Sparse SS2D · eTram | mAP: 30.2 -> 32.6 | Helped |
| add the IPL-Scan local constraint (ablation) | SMamba Sparse SS2D · eTram | mAP: 30.7 -> 32.6 | Helped |
| add Bidi-channel Scan in place of MLP (ablation) | SMamba GCI · eTram | mAP: 31.0 -> 31.5 | Helped |
| add 1 × 1 convolution alongside Bidi-channel Scan (ablation) | SMamba GCI · eTram | mAP: 31.5 -> 32.6 | Helped |
| place GCI after the final two stages (S3–S4), rather than after S4 alone (ablation) | SMamba · eTram | mAP: 31.0 -> 32.6 | Helped |
| apply the proposed sparsification strategy | SMamba baseline · Gen1 | mAP: 50.0 -> 50.4 | Helped |
| apply the proposed sparsification strategy | SMamba baseline · 1Mpx | mAP: 48.8 -> 49.3 | Helped |
| apply the proposed sparsification strategy | SMamba baseline · eTram | mAP: 32.3 -> 32.6 | Helped |
Surprising findings
- Variance and entropy scoring underperform the learned scoring module on eTram, and the learnable scoring module itself is suboptimal due to indirect gradient optimization.
Limitations
- STCA assumes temporal and spatial independence of noise, which may fail for large noise bursts or clusters.
- Sparsity scaling factor beta is manually adjusted rather than learned.
Experimental setup
Datasets: Gen1 1Mpx eTram
Benchmarks: Gen1 1Mpx eTram
Metrics: COCO mAP FLOPs parameters runtime
Baselines: RED ASTMNet ERGO-12 RVT-B GET-T SAST-CB S5-ViT-B VSS
Builds on: Mamba VMamba RVT SAST YOLOX
Closest papers by meaning
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · similarity 0.88)
- EventMamba: Enhancing Spatio-Temporal Locality with State Space Models for Event-Based Video Reconstruction (AAAI 2025 · similarity 0.88)
- Scene Adaptive Sparse Transformer for Event-based Object Detection (CVPR 2024 · similarity 0.88)
- 3DET-Mamba: Causal Sequence Modelling for End-to-End 3D Object Detection (NeurIPS 2024 · similarity 0.86)
- Samba: A Unified Mamba-based Framework for General Salient Object Detection (CVPR 2025 · similarity 0.85)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 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.