Research map / Label-efficient and robust learning / OOD and anomaly detection
Entropy Maximization and Meta Classification for Out-of-Distribution Detection in Semantic Segmentation
Robin Chan, Matthias Rottmann, Hanno Gottschalk
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 |
|---|---|---|---|
| add OoD training with entropy maximization on COCO (ablation) | DeepLabv3+ WideResNet38 baseline · LostAndFound pixel-wise | AUPRC: 0.46 -> 0.76 | Improved |
| add OoD training with entropy maximization on COCO (ablation) | DeepLabv3+ WideResNet38 baseline · Fishyscapes Val pixel-wise | AUPRC: 0.28 -> 0.81 | Improved |
| add OoD training with entropy maximization (ablation) | DualGCNNet · LostAndFound pixel-wise | AUPRC: 0.36 -> 0.51 | Improved |
| add OoD training with entropy maximization (ablation) | DualGCNNet · Fishyscapes Val pixel-wise | AUPRC: 0.07 -> 0.38 | Improved |
| add OoD training with entropy maximization (ablation) | DeepLabv3+ WideResNet38 · LostAndFound pixel-wise | FPR95: 0.35 -> 0.09 | Improved |
| add OoD training with entropy maximization (ablation) | DeepLabv3+ WideResNet38 · Cityscapes val | mIoU: 0.90 -> 0.89 | Hurt |
| add OoD training with entropy maximization (ablation) | DualGCNNet · Cityscapes val | mIoU: 0.80 -> 0.76 | Hurt |
| lower entropy threshold to t=0.2 for OoD prediction (ablation) | DeepLabv3+ WideResNet38 baseline · Cityscapes val | mIoU: 0.05 baseline and 0.03 re-trained mIoU decrease | Hurt |
| add a meta classifier on top of entropy thresholding (ablation) | OoD-trained model at t=0.3 · LostAndFound segment-wise | F1: 0.26 -> 0.82 | Improved |
| add a meta classifier on top of entropy thresholding (ablation) | baseline model at t=0.3 · LostAndFound segment-wise | F1: 0.28 -> 0.79 | Improved |
| add OoD training (ablation) | baseline + meta classifier at t=0.7 · LostAndFound meta classification | AUPRC: 0.5283 -> 0.7185 | Improved |
| add OoD training (ablation) | baseline + meta classifier at t=0.1 · LostAndFound meta classification | AUPRC: 0.9993 -> 0.9993 | No clear effect |
Closest papers by meaning
- MOS: Towards Scaling Out-of-Distribution Detection for Large Semantic Space (CVPR 2021 · similarity 0.90)
- Deep Hybrid Models for Out-of-Distribution Detection (CVPR 2022 · similarity 0.90)
- Triggering Failures: Out-of-Distribution Detection by Learning From Local Adversarial Attacks in Semantic Segmentation (ICCV 2021 · similarity 0.89)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.89)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.88)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.88)
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.
