Research map / Label-efficient and robust learning / OOD and anomaly detection
Revisiting Out-of-Distribution Detection in Real-Time Object Detection: From Benchmark Pitfalls to a New Mitigation Paradigm
Changshun Wu, Weicheng He, Chih-Hong Cheng, Xiaowei Huang, Saddek Bensalem
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on proximal OoD samples treated as background | YOLOv10 · BDD-100K Near-OoD | OoD detection count: 701 -> 133 | Improved |
| add KNN OoD filter to the fine-tuned detector | fine-tuned YOLOv10 · BDD-100K Near-OoD | OoD detection count: 133 -> 80 | Improved |
| fine-tune on proximal OoD samples | Faster R-CNN · BDD-100K Near-OoD | OoD detection count: 2576 -> 394 | Improved |
| fine-tune on proximal OoD samples | RT-DETR · BDD-100K Near-OoD | OoD detection count: 3145 -> 967 | Improved |
| add KNN filter to fine-tuned RT-DETR | fine-tuned RT-DETR · PASCAL-VOC Far-OoD | OoD detection count: 966 -> 470 | Improved |
| discard proposals whose background score is highest at inference (ablation) | Faster R-CNN · PASCAL-VOC | mAP: 0.402 -> 0.391 | Mixed |
| remove ID outlier objects from calibration/test ID data (ablation) | KNN filter · PASCAL-VOC vs OpenImages | FPR95: 0.5376 -> 0.4467 | Improved |
| remove ID-labelled objects from the OoD test set (fix Type 1 error) (ablation) | OoD benchmark evaluation · PASCAL-VOC vs OpenImages | FPR95: +9.85% inflation removed | Improved |
| use error-prone ID subset (2,000 images) instead of all ID data (ablation) | proximal OoD fine-tuning · BDD-100K | mAP / Near-OoD count: 0.2556 -> 0.2492; 133 -> 170 | Mixed |
| use Object365 instead of Caltech256 as proximal OoD source (ablation) | proximal OoD fine-tuning (ID_err) · BDD-100K | Near-OoD count: 170 -> 118 | Improved |
| select proximal categories by closest CLIP text similarity instead of farthest (ablation) | proximal OoD fine-tuning · BDD-100K | Near-OoD count: 226 -> 143 | Improved |
| closer proximal OoD data (ablation) | proximal OoD fine-tuning · BDD-100K | mAP (all): 0.2523 -> 0.2457 | Hurt |
Closest papers by meaning
- RUNA: Object-Level Out-of-Distribution Detection via Regional Uncertainty Alignment of Multimodal Representations (AAAI 2025 · similarity 0.91)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.90)
- Revisiting Logit Distributions for Reliable Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.90)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.90)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.89)
- Fast Decision Boundary based Out-of-Distribution Detector (ICML 2024 · similarity 0.89)
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.
