Research map / Detection and segmentation / Open-set and few-shot detection
Towards Evidential and Class Separable Open Set 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 EOD, an Evidential Object Detector that replaces the Faster R-CNN classification head with an evidential deep learning head, trained with a hybrid evidential loss, weak regularization, and contrastive learning, to detect and reject unknown objects in open set detection.
Problem
Open set object detection: closed-set detectors overconfidently misclassify unknown objects as known classes, so they need reliable unknown identification without losing known-class accuracy.
Approach
Formulates OSOD as uncertainty estimation via evidential deep learning (Dirichlet prior over class probabilities) plugged in as a replacement classification head, with a hybrid CE/MSE Bayes-risk loss, a weak KL regularizer that penalizes only excessive evidence for wrong labels, contrastive class separation, and joint foreground/background uncertainty thresholds for unknown identification.
Key results
- On VOC-COCO-20, EOD reaches WI 11.57, AOSE 8989, mAPK 58.97 and APu 16.55, versus OpenDet's 14.58/17754/57.72 and APu 14.07 in the Table 1 comparison (WI and AOSE values are from the table rows for OpenDet at VOC-COCO-20).
- On VOC-COCO-40 and -60, EOD achieves the highest APu (11.53 and 4.75) and lowest AOSE among compared methods.
- Closed-set mAPK on VOC is 79.96 for EOD vs 80.10 for closed-set Faster R-CNN, i.e., comparable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the hybrid evidential loss with CE Bayes-risk loss (ablation) | EOD · VOC-COCO-20 | mAPK: 58.97 -> 48.72 | Hurt |
| replace the hybrid evidential loss with MSE Bayes-risk loss (ablation) | EOD · VOC-COCO-20 | AOSE: 8989 -> 13064 | Hurt |
| add weak regularization instead of no regularization (ablation) | EOD · VOC-COCO-20 | mAPK: 58.45 -> 58.97 | Helped |
| replace weak regularization with original KL regularization (ablation) | EOD · VOC-COCO-20 | mAPK: 58.97 -> 56.90 | Hurt |
| add auxiliary contrastive learning (ablation) | EOD · VOC-COCO-20 | AOSE: 10515 -> 8989 | Helped |
| model background uncertainty in addition to foreground uncertainty (ablation) | EOD · VOC-COCO-20 | APu: 12.68 -> 14.40 | Helped |
| use a lower uncertainty threshold for background proposals (ablation) | EOD · VOC-COCO-20 | APu: 14.40 -> 16.55 | Helped |
| use EOD instead of OpenDet | Faster R-CNN, ResNet-50 FPN · VOC-COCO-20 | mAPK: 57.72 -> 58.97 | Helped |
| use EOD instead of OpenDet | Faster R-CNN, ResNet-50 FPN · VOC-COCO-20 | APu: 14.07 -> 16.55 | Helped |
| use EOD instead of Faster R-CNN | Faster R-CNN, ResNet-50 FPN · VOC | mAPK: 80.10 -> 79.96 | Hurt |
| use EOD instead of Faster R-CNN | Faster R-CNN, ResNet-50 FPN · VOC-COCO-5000 | mAPK: 74.52 -> 74.38 | Hurt |
| use EOD instead of OpenDet | Faster R-CNN, ResNet-50 FPN · VOC-COCO-20000 | APu: 16.54 -> 16.76 | Helped |
Surprising findings
- The original EDL KL regularizer degrades performance in detection (mAPK drops from 58.45 to 56.90), while the proposed weak regularizer helps.
- Modeling background-proposal uncertainty with a lower threshold substantially raises APu (14.40 to 16.55) without affecting WI, AOSE or mAPK.
Limitations
- Relies on hand-tuned uncertainty thresholds (tau_f=0.02, tau_b set to half of tau_f) and annealing schedules.
- On low-openness and some extreme settings the closed-set Faster R-CNN retains the highest mAPK.
Experimental setup
Datasets: PASCAL VOC MS COCO
Benchmarks: PASCAL VOC VOC-COCO-20 VOC-COCO-40 VOC-COCO-60 VOC-COCO-2500 VOC-COCO-5000 VOC-COCO-20000
Metrics: WI AOSE mAPK APu
Baselines: Faster R-CNN PROSER ORE DS OpenDet
Builds on: Evidential Deep Learning Faster R-CNN OpenDet Feature Pyramid Network
Closest papers by meaning
- Meta Evidential Transformer for Few-Shot Open-Set Recognition (ICML 2024 · similarity 0.87)
- Open-World Objectness Modeling Unifies Novel Object Detection (CVPR 2025 · similarity 0.86)
- ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning (AAAI 2024 · similarity 0.86)
- Looking Beyond the Known: Towards a Data Discovery Guided Open-World Object Detection (NeurIPS 2025 · similarity 0.85)
- Hyp-OW: Exploiting Hierarchical Structure Learning with Hyperbolic Distance Enhances Open World Object Detection (AAAI 2024 · similarity 0.85)
- RUNA: Object-Level Out-of-Distribution Detection via Regional Uncertainty Alignment of Multimodal Representations (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.