Research map / Image generation / Autoregressive and latent generation
Self-Feature Distillation with Uncertainty Modeling for Degraded Image Recognition
Zhou Yang, Weisheng Dong, Xin Li, Jinjian Wu, Leida Li, Guangming Shi
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 self-feature distillation to cross-entropy fine-tuning (ablation) | ResNet50 robust model · ImageNet-C | mCE: 51.31% -> 49.50% | Improved |
| add uncertainty modeling to self-feature distillation (ablation) | ResNet50 robust model · ImageNet-C | mCE: 49.50% -> 46.37% | Improved |
| add uncertainty estimation module (ablation) | ResNet50 · ImageNet-C unseen corruptions | top-1 accuracy: 60.23% -> 62.90% | Improved |
| add uncertainty estimation module (ablation) | ResNeXt101 · ImageNet-C unseen corruptions | top-1 accuracy: 65.30% -> 67.83% | Improved |
| add uncertainty estimation module (ablation) | ResNeXt101 · ImageNet-C | mCE: 41.23% -> 39.50% | Improved |
| use spatial-wise instead of sample-wise uncertainty (ablation) | Ours · ImageNet-C | mCE: 46.93% -> 46.37% | Improved |
| set lambda to 0.1 instead of 0 (ablation) | Ours · ImageNet-C | top-1 accuracy: 58.43% -> 63.44% | Improved |
| raise lambda from 0.1 to 1 (ablation) | Ours · ImageNet-C | top-1 accuracy: 63.44% -> 62.78% | Hurt |
| reconstruct deep semantic features instead of shallow features (ablation) | Ours · ImageNet-C | top-1 accuracy: 4.5 % higher | Improved |
| add the framework to DeepAugment training | DeepAugment · ImageNet-C | mCE: 60.31% -> 59.04% | Improved |
| add the framework to Augmix training | Augmix · ImageNet-C | mCE: 65.30% -> 64.37% | Improved |
| add the framework to DeepAugment+Augmix | DeepAugment+Augmix · HQ ImageNet | top-1 accuracy: 73.64% -> 73.81% | Improved |
Closest papers by meaning
- Self Supervision to Distillation for Long-Tailed Visual Recognition (ICCV 2021 · similarity 0.86)
- Vector Quantization With Self-Attention for Quality-Independent Representation Learning (CVPR 2023 · similarity 0.85)
- Feature Normalized Knowledge Distillation for Image Classification (ECCV 2020 · similarity 0.85)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.84)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (AAAI 2024 · similarity 0.84)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.83)
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.
