Research map / Detection and segmentation / Open-set and few-shot detection
YolOOD: Utilizing Object Detection Concepts for Multi-Label Out-of-Distribution 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.
Converts a YOLOv5s object detector into a multi-label image classifier whose objectness-times-class scores serve as an OOD score, trained on Grounding DINO auto-annotated boxes, and introduces Objects365in/out and NUS-WIDEout benchmarks.
Problem
Out-of-distribution detection has been studied mainly for multi-class classification, while the multi-label setting, which is more common in practice, remains underexplored with few methods and oversimplified benchmarks.
Approach
Replace YOLO's detection heads with a simplified layer predicting objectness and class scores per cell, enlarge the responsible cells per object via a fraction p_k, and use the max over candidates of objectness times class score (aggregated over heads) as the OOD score, so object detectors' inherent separation of objects of interest from irrelevant objects gives OOD detection without external OOD
Key results
- On Objects365out with PASCAL-VOC in-distribution, YolOOD reduces FPR95 by 9.53% relative to the best baseline.
- On MS-COCO in-distribution with Objects365out OOD, YolOOD reduces FPR95 by 3.10% relative to the best baseline.
- Objects365in in-distribution with NUS-WIDEout OOD: FPR95 decreases by 12.27% relative to other OOD methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · PASCAL-VOC → Objects365out | FPR95: 27.90 -> 16.38 | Helped |
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · MS-COCO → Objects365out | FPR95: 14.80 -> 11.53 | Helped |
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · Objects365in → Objects365out | FPR95: 23.13 -> 17.24 | Helped |
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · PASCAL-VOC → NUS-WIDEout | FPR95: 20.19 -> 18.47 | Helped |
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · MS-COCO → NUS-WIDEout | FPR95: 8.29 -> 4.40 | Helped |
| Use YolOOD instead of JointEnergy | YOLOv5s multi-label OOD detection · Objects365in → NUS-WIDEout | FPR95: 24.46 -> 9.54 | Helped |
| Use the joint objectness-class score rather than either score alone (ablation) | YolOOD · OOD detection benchmarks | OOD detection performance | Helped |
| Set responsible-cell proportions to (p1, p2, p3) = (0.0, 0.1, 0.5) (ablation) | YolOOD · PASCAL-VOC, MS-COCO, and Objects365in | in-distribution mAP | Helped |
| Sum candidate-set scores and take the maximum over class scores instead of alternative aggregation functions (ablation) | YolOOD · OOD detection benchmarks | FPR95 | Mixed |
| Apply JointEnergy to YolOOD scores instead of the regular YolOOD score (ablation) | YolOOD · MS-COCO → Objects365out | FPR95: 11.53% -> 7.1% | Helped |
| Use YolOOD instead of vanilla YOLO (ablation) | Object-detection-based OOD detection · PASCAL-VOC → Objects365out | FPR95: 40.57% -> 16.38% | Helped |
Surprising findings
- The vanilla YOLO detector performs worse than YolOOD at OOD detection, with the objectness score alone being a strong OOD signal in some settings.
- Mahalanobis performs poorly on the multi-label setting, far below logit-based baselines.
Limitations
- Requires bounding box annotations, which are obtained here via Grounding DINO auto-labeling from class names.
- Evaluated only on image classification benchmarks with YOLOv5s backbone.
- Responsible-cell fraction p_k selected empirically on the in-distribution datasets.
Experimental setup
Datasets: PASCAL VOC MS-COCO Objects365 NUS-WIDE ImageNet-22K Textures
Benchmarks: PASCAL-VOC vs Objects365out MS-COCO vs Objects365out Objects365in vs Objects365out PASCAL-VOC vs NUS-WIDEout MS-COCO vs NUS-WIDEout Objects365in vs NUS-WIDEout
Metrics: FPR95 AUROC AUPR
Baselines: MaxLogit MSP ODIN Mahalanobis JointEnergy YOLO-cls vanilla YOLO
Builds on: YOLOv5 MaxLogit JointEnergy Grounding DINO
Closest papers by meaning
- A Noisy Elephant in the Room: Is Your Out-of-Distribution Detector Robust to Label Noise? (CVPR 2024 · similarity 0.89)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.89)
- EAT: Towards Long-Tailed Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- MultiOOD: Scaling Out-of-Distribution Detection for Multiple Modalities (NeurIPS 2024 · similarity 0.88)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.87)
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.87)
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.