Research map / Detection and segmentation / Weakly supervised detection and segmentation
LEOD: Label-Efficient Object Detection for Event Cameras
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 LEOD, a two-stage offline self-training framework for label-efficient event-based object detection that uses time-flip TTA, tracking-based filtering, and soft anchor assignment on RVT detectors.
Problem
Event-based object detection requires costly dense labeling of high-temporal-resolution event streams; existing methods train only on labeled events and discard unlabeled data, so label-efficient learning under weakly- and semi-supervised settings is needed.
Approach
Pre-train a detector on limited labels, generate pseudo labels offline using bi-directional (time-flip) test-time augmentation and forward-backward tracking, then re-train with soft anchor assignment that ignores loss on low-confidence pseudo boxes.
Key results
- On Gen1 WSOD, LEOD improves RVT-S mAP by 8.6% and 7.8% with 1% and 2% labels (39.4% and 42.2%).
- On 1Mpx, RVT-S with 10% labels (44.6%) surpasses RVT-S trained on 100% labels (44.1%).
- With all labels, LEOD-RVT-S reaches 48.7% mAP on Gen1 vs 46.5% for RVT-S, a new state of the art among models without pretraining.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply LEOD self-training with TTA, tracking, and soft anchor assignment | RVT-S · Gen1 WSOD 1% labels | mAP (%): 30.8 -> 39.4; +8.6 | Helped |
| Apply LEOD self-training with TTA, tracking, and soft anchor assignment | RVT-S · Gen1 WSOD 100% labels | mAP (%): 46.5 -> 48.7; +2.2 | Helped |
| Apply LEOD self-training with TTA, tracking, and soft anchor assignment | RVT-S · 1Mpx, all available labels | mAP (%): 44.1 -> 46.7; +2.6 | Helped |
| Apply LEOD self-training with TTA, tracking, and soft anchor assignment | RVT-B · Gen1 WSOD 1% labels | mAP (%): 31.6 -> 40.0; +8.4 | Helped |
| Add time-flip and horizontal-flip test-time augmentation (ablation) | Vanilla Self-Training with RVT-S · Gen1 WSOD 1% labels | mAP (%): 35.2 -> 38.2 | Helped |
| Add tracking-based post-processing (ablation) | RVT-S with TTA · Gen1 WSOD 1% labels | mAP (%): 38.2 -> 38.5 | Helped |
| Add soft anchor assignment (ablation) | RVT-S with TTA and tracking · Gen1 WSOD 1% labels | mAP (%): 38.5 -> 39.4 | Helped |
| Add time-flip and horizontal-flip test-time augmentation (ablation) | Vanilla Self-Training with RVT-S · 1Mpx WSOD 1% labels | mAP (%): 33.5 -> 37.3 | Helped |
| Add tracking-based post-processing (ablation) | RVT-S with TTA · 1Mpx WSOD 1% labels | mAP (%): 37.3 -> 37.5 | Helped |
| Add soft anchor assignment (ablation) | RVT-S with TTA and tracking · 1Mpx WSOD 1% labels | mAP (%): 37.5 -> 38.4 | Helped |
| Vary hard and soft confidence thresholds (ablation) | LEOD-RVT-S · Gen1 WSOD | mAP (%): 38.9 to 39.4 (1%); 44.9 to 45.6 (10%) | No clear effect |
| Remove soft anchor assignment and use a high filtering threshold (ablation) | LEOD-RVT-S · Gen1 WSOD 1% labels | mAP (%): 39.4 -> 38.4 | Hurt |
Surprising findings
- On 1Mpx, pseudo-labeling baselines perform worse than the supervised pre-trained RVT-S at 5% and 10% labels, showing the importance of label screening.
- Naive self-training with a high confidence threshold (0.9, common in prior work) yields clearly lower mAP than low thresholds with soft anchor assignment.
Limitations
- Only intra-dataset experiments were conducted; no cross-dataset training or generalization evaluation.
- Pseudo-label quality depends on pre-trained teacher; failure cases are discussed in the appendix.
- Hyperparameters such as thresholds and minimum track length are set per category and dataset by rule-based choices.
Experimental setup
Datasets: Gen1 1Mpx
Benchmarks: Gen1 1Mpx
Metrics: mAP
Baselines: Supervised Baseline Vanilla Self-Training Soft Teacher RED ASTMNet HMNet-L3 ERGO-12 RVT-B
Builds on: RVT Soft Teacher YOLOX
Closest papers by meaning
- Learning to Detect Objects with a 1 Megapixel Event Camera (NeurIPS 2020 · similarity 0.89)
- Unbiased Teacher v2: Semi-Supervised Object Detection for Anchor-Free and Anchor-Based Detectors (CVPR 2022 · similarity 0.89)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.87)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.86)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.86)
- Object Detection using Event Camera: A MoE Heat Conduction based Detector and A New Benchmark Dataset (CVPR 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.