Research map / Label-efficient and robust learning / Recognition and long-tail
BackMix: Regularizing Open Set Recognition by Removing Underlying Fore-Background Priors
Yu Wang, Junxian Mu, Hongzhi Huang, Qilong Wang, Pengfei Zhu, Qinghua Hu
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 BackMix | Plain* softmax baseline · CIFAR10 (6 known/4 unknown) | AUROC: 67.7 -> 91.3 | Improved |
| add BackMix | ARPL · CIFAR10 (6 known/4 unknown) | AUROC: 89.8 -> 91.0 | Improved |
| add BackMix | ARPL · Tiny-ImageNet | AUROC: 76.0 -> 76.3 | No clear effect |
| add BackMix | CSSR · CIFAR10 (6 known/4 unknown) | AUROC: 90.7 -> 94.2 | Improved |
| replace BackMix with Cutmix | Plain* softmax · ImageNet30 split | AUIN: 86.4 -> 76.7 | Hurt |
| replace BackMix with Mixup | Plain* softmax · ImageNet30 split | AUIN: 86.4 -> 80.8 | Hurt |
| add Cutout | Plain* softmax · ImageNet30 split | AUROC: 89.9 -> 90.4 | No clear effect |
| use backgrounds from any random image instead of only different-class images (ablation) | BackMix · CIFAR10 (6 known/4 unknown) | AUROC: 90.9 -> 91.3 | No clear effect |
| train on FG+BG* (foreground with random other backgrounds) instead of raw images (ablation) | COCO 12-class ResNet18 · COCO known/raw test | AUROC: 67.2 -> 69.5 | Improved |
| train on FG only instead of raw images (ablation) | COCO 12-class ResNet18 · COCO known/raw test | AUROC: 67.2 -> 52.7 | Hurt |
| train on FG+BG* instead of raw, test with backgrounds replaced (ablation) | COCO 12-class ResNet18 · COCO FG+BG*/Raw test | AUROC: 59.0 -> 83.6 | Improved |
| add outlier exposure with DTD outliers (ablation) | Plain WRN-40-4 · CIFAR10 known | AUROC: 87.9 -> 77.4 | Hurt |
Closest papers by meaning
- Towards Accurate Open-Set Recognition via Background-Class Regularization (ECCV 2022 · similarity 0.89)
- The Devil is in the Wrongly-classified Samples: Towards Unified Open-set Recognition (ICLR 2023 · similarity 0.87)
- Frequency Shuffling and Enhancement for Open Set Recognition (AAAI 2024 · similarity 0.85)
- Exploring Diverse Representations for Open Set Recognition (AAAI 2024 · similarity 0.85)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.84)
- All Beings Are Equal in Open Set Recognition (AAAI 2024 · similarity 0.84)
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.
