Research map / Label-efficient and robust learning / Recognition and long-tail
Class-Specific Semantic Reconstruction for Open Set Recognition
Hongzhi Huang, Yu Wang, Qinghua Hu, Ming‐Ming Cheng
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 |
|---|---|---|---|
| use pixelwise softmax then average pooling instead of average pooling then softmax (ablation) | CSSR · CIFAR10 unknown detection average | AUROC: 94.7 -> 97.4 | Improved |
| use MAE instead of MSE as reconstruction error (ablation) | CSSR · CIFAR10 unknown detection average | AUROC: 96.9 -> 97.4 | Improved |
| replace linear classification layer with CSSR class-specific autoencoders (ablation) | linear baseline · CIFAR10 unknown detection average | AUROC: 94.4 -> 97.4 | Improved |
| add pixelwise softmax strategy to linear classifier (ablation) | Linear · CIFAR10 unknown detection average | AUROC: 94.4 -> 94.7 | Improved |
| replace CSSR with RCSSR on CIFAR10 unknown detection (ablation) | CSSR · CIFAR10 unknown detection average | AUROC: 97.4 -> 97.3 | No clear effect |
| integrate RRE and feature magnitude into score for CSSR vs max softmax-like s1 (ablation) | CSSR · ImageNet-30 | AUROC: 92.2 -> 94.8 | Improved |
| use fused score s_all instead of best individual score (ablation) | CSSR · ImageNet-30 | AUROC: 95.5 -> 95.3 | No clear effect |
| add training-time CSSR layer vs baseline on CIFAR10 closed set | baseline · CIFAR10 closed set | accuracy: 95.1 -> 95.3 | Improved |
| replace CSSR with RCSSR on closed set | CSSR · CIFAR10 closed set | accuracy: 95.3 -> 95.6 | Improved |
| use CSSR instead of plain baseline on ImageNet-1000 split | baseline · ImageNet-1000 split | TNR@TPR95: 51.8 -> 62.4 | Improved |
| use RCSSR instead of CSSR on large-scale iNaturalist unknowns | CSSR · iNaturalist unknown | AUROC: 93.8 -> 94.5 | Improved |
| use CSSR instead of ARPL on CIFAR+10 | ARPL · CIFAR+10 unknown detection | AUROC: 97.1 -> 96.3 | Hurt |
Closest papers by meaning
- Classification-Reconstruction Learning for Open-Set Recognition (CVPR 2019 · similarity 0.89)
- Generative-Discriminative Feature Representations for Open-Set Recognition (CVPR 2020 · similarity 0.89)
- Open-Set Recognition: A Good Closed-Set Classifier is All You Need (ICLR 2022 · similarity 0.88)
- All Beings Are Equal in Open Set Recognition (AAAI 2024 · similarity 0.87)
- Unlocking Better Closed-Set Alignment Based on Neural Collapse for Open-Set Recognition (AAAI 2025 · similarity 0.86)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 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.
