Research map / Label-efficient and robust learning / Recognition and long-tail
All Beings Are Equal in Open Set Recognition
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 DCTAU, which replaces the single K+1 pseudo-unknown class with K Target-Aware Universum classes generated by targeted mixup and trains with a dual contrastive loss for open-set recognition.
Problem
Open-set recognition typically models all unknowns as one pseudo-unknown (K+1) class, which distorts unknown class distributions and causes class- and instance-wise imbalance between known and unknown data.
Approach
Generate K target-aware universum (TAU) classes via Targeted Mixup of each known class with the average of the other known classes, and apply a Dual Contrastive loss that treats TAU classes symmetrically with known classes, yielding a K+K strategy.
Key results
- DCTAU reaches 83.6% AUROC on TinyImageNet, versus 80.9% for ConOSR (the prior best contrastive method).
- DCTAU reaches 95.6% AUROC on CIFAR10 and 99.2% on SVHN in the open-set recognition table.
- On OSCR, DCTAU improves TinyImageNet by about 11.7% over the second-best method, per the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DCTAU instead of ConOSR | OSR models · SVHN | AUROC: 99.1 -> 99.2 | Helped |
| use DCTAU instead of OpenHybrid | OSR models · CIFAR10 | AUROC: 95.0 -> 95.6 | Helped |
| use DCTAU instead of All-U-Need(MLS) | OSR models · TinyImageNet | AUROC: 83.0 -> 83.6 | Helped |
| use DCTAU instead of ARPL+CS | OSR models · TinyImageNet | OSCR: 65.9 -> 77.6 | Helped |
| add the TAU-side DC loss (ablation) | DCTAU · CIFAR10 | AUROC: 92.3 -> 95.6 | Helped |
| add the TAU-side DC loss (ablation) | DCTAU · TinyImageNet | AUROC: 71.0 -> 83.6 | Helped |
| replace TAU with traditional Mixup images (ablation) | DCTAU pseudo-unknown generation · CIFAR10 known classes | AUROC: 95.6 -> 93.8 | Hurt |
| replace TAU with traditional Mixup images (ablation) | DCTAU pseudo-unknown generation · CIFAR10 known classes | OSCR: 93.9 -> 92.2 | Hurt |
| replace TAU with Noise-Gaussian pseudo-unknowns (ablation) | DCTAU pseudo-unknown generation · CIFAR10 known classes | AUROC: 95.6 -> 83.2 | Hurt |
| replace TAU with Noise-Uniform pseudo-unknowns (ablation) | DCTAU pseudo-unknown generation · CIFAR10 known classes | AUROC: 95.6 -> 85.1 | Hurt |
| replace TAU with Natural Images pseudo-unknowns (ablation) | DCTAU pseudo-unknown generation · CIFAR10 known classes | AUROC: 95.6 -> 91.1 | Hurt |
| set TAU mixup weight λ to 0.5 (ablation) | DCTAU · CIFAR10 known classes | AUROC: 93.5, 94.2, 95.6, 95.1, 95.0 | Helped |
Surprising findings
- Removing the dual contrastive loss (DCTAU w/o DC) causes a significant drop, to 71.0% AUROC on TinyImageNet, even below several baselines.
Limitations
- Theoretical analysis is limited to gradient analysis of the loss, with the authors noting further theoretical foundations are needed.
- The K+K choice is a compromise justified empirically on partial CIFAR10 only.
Experimental setup
Datasets: MNIST SVHN CIFAR10 CIFAR100 TinyImageNet Omniglot ImageNet LSUN
Benchmarks: MNIST SVHN CIFAR10 CIFAR+10 CIFAR+50 TinyImageNet Omniglot MNIST-Noise Noise ImageNet LSUN
Metrics: AUROC OSCR macro F1-score
Baselines: Softmax OpenMax G-OpenMax OSRCI CPN RPL++ GFROSR PROSER OpenHybrid ARPL Class-inclusion PMAL
Builds on: SupCon Universum learning Mixup ConOSR ARPL
Closest papers by meaning
- Learning Network Architecture for Open-Set Recognition (AAAI 2022 · similarity 0.90)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.87)
- Frequency Shuffling and Enhancement for Open Set Recognition (AAAI 2024 · similarity 0.87)
- Exploring Diverse Representations for Open Set Recognition (AAAI 2024 · similarity 0.86)
- Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · similarity 0.85)
- Towards Evidential and Class Separable Open Set Object Detection (AAAI 2024 · 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.