Research map / Detection and segmentation / CNN features
FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies
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 FlexEvent, an event-frame object detector (FlexFuse adaptive fusion plus FlexTune frequency-adaptive self-training) that keeps detection accuracy across operational frequencies from 20 Hz to 180 Hz on DSEC benchmarks.
Problem
Event-based object detectors are tied to fixed frame-rate temporal windows and degrade sharply at high operational frequencies, where event data is sparse and lacks semantic detail, while labels exist only at low frame rates.
Approach
Fuse high-frequency event features with RGB frame features through learned, noise-perturbed soft gating (FlexFuse), and generate pseudo-labels for unlabeled high-frequency event data via bootstrapping, temporal consistency filtering (bidirectional augmentation, NMS, IoU tracklet pruning), and cyclic self-training (FlexTune).
Key results
- On DSEC-Det validation, FlexEvent reaches 57.4 mAP versus 46.7 for the best prior fusion method HDI-Former and 41.9 for DAGr-50.
- Full FlexEvent keeps 57.4% mAP at 20 Hz and 50.9% at 180 Hz, versus 22.9% at 180 Hz for event-only RVT-style baseline without FlexTune or FlexFuse.
- On DSEC-Detection, FlexEvent averages 47.4 mAP versus 38.0 for CAFR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add event-frame fusion with FlexFuse (ablation) | FlexEvent · DSEC-Det, average across frequencies | mAP: 43.7% to 56.4% | Helped |
| add FlexTune frequency-adaptive self-training (ablation) | FlexEvent with FlexFuse · DSEC-Det, 180 Hz | mAP: 49.2% -> 50.9% | Helped |
| add FlexTune frequency-adaptive self-training (ablation) | event-only detector · DSEC-Det, 180 Hz | mAP: 22.9% to 30.4% | Helped |
| use interpolated high-frequency labels instead of FlexTune (ablation) | FlexFuse · DSEC-Det, 180 Hz | mAP: 50.9% -> 48.3% | Hurt |
| add FlexFuse event-frame fusion (ablation) | FlexTune-trained detector · DSEC-Det, 180 Hz | mAP: 30.4% -> 50.9% | Helped |
| operate at 90 Hz instead of 20 Hz | FlexEvent · DSEC-Det | performance retained: 96.2% | Hurt |
Surprising findings
- Interpolated ground-truth high-frequency labels underperform the FlexTune pseudo-labels, with 48.3% vs 50.9% at 180 Hz in the ablation.
- Event-only models without FlexTune lose more than half their mAP from 20 Hz to 180 Hz (53.2% to 22.9%).
Limitations
- Model has 45.4M parameters, more than event-only baselines, and carries the full RGB branch cost.
- Evaluated only on DSEC driving datasets with Car and Pedestrian classes emphasized.
- Fusion requires temporally aligned RGB frames, so it does not handle event-only deployments.
Experimental setup
Datasets: DSEC-Det DSEC-Detection DSEC-MOD
Benchmarks: DSEC-Det DSEC-Detection DSEC-MOD
Metrics: mAP AP50 AP75 AP_S AP_M AP_L inference time (ms)
Baselines: RVT SAST SSM LEOD DAGr-50 HDI-Former CAFR RENet
Builds on: RVT DAGr YOLO-X ResNet-50
Closest papers by meaning
- EGSST: Event-based Graph Spatiotemporal Sensitive Transformer for Object Detection (NeurIPS 2024 · similarity 0.88)
- LEOD: Label-Efficient Object Detection for Event Cameras (CVPR 2024 · similarity 0.87)
- PASS: Path-selective State Space Model for Event-based Recognition (NeurIPS 2025 · similarity 0.87)
- EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision (AAAI 2025 · similarity 0.87)
- Object Detection using Event Camera: A MoE Heat Conduction based Detector and A New Benchmark Dataset (CVPR 2025 · similarity 0.87)
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.86)
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.