Research map / Label-efficient and robust learning / Recognition and long-tail
Distance Based Image Classification: A solution to generative classification’s conundrum?
Wen-Yan Lin, Siying Liu, Bing Tian Dai, Hongdong Li
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 |
|---|---|---|---|
| replace Soft-Max/discriminative layer with strictly incremental distance classifier | ResNet-50 features · STL-10 | classification accuracy: 0.967 -> 0.970 | Improved |
| replace kernel SVM with distance classifier | ResNet-50 features / pixels · MNIST | classification accuracy: 0.966 -> 0.955 | Hurt |
| replace kernel SVM with distance classifier | pixels · Fashion-MNIST | classification accuracy: 0.884 -> 0.825 | Hurt |
| use distance classifier instead of Shell-R strict incremental learner | strict incremental learners · STL-10 | classification accuracy: 0.930 -> 0.970 | Improved |
| use distance classification vs best replay-based baseline | incremental learners · SUN-397 (top-5) | final accuracy: 72.6 -> 87.5 | Improved |
| use distance classification vs replay-based baselines | incremental learners · ImageNet-1000 (top-5) | final accuracy: 73.2 -> 88.0 | Improved |
| evaluate with corrected ImageNet labels | Distance classification · ImageNet-1000 (top-5) | final accuracy: 88.0 -> 91.7 | Improved |
| use distance classification vs iTAML | incremental learners · Food-101 (top-5) | final accuracy: 84.7 -> 81.1 | Hurt |
| use distance classification vs REMIND | incremental learners · SUN-397 incremental update | training time for last increment (s): 2452 -> 453 | Improved |
| use distance classification vs REMIND | incremental learners · Flower-102 incremental update | training time for last increment (s): 19 -> 38 | Hurt |
| add normalization to Euclidean nearest neighbor (ablation) | Euc. NN · STL-10 | validation AUPRC: 0.764 -> 0.922 | Improved |
| use distance classifier instead of raw Euclidean NN (ablation) | strict incremental · STL-10 | validation AUPRC: 0.764 -> 0.984 | Improved |
Closest papers by meaning
- Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.89)
- Generative Classifiers Avoid Shortcut Solutions (ICLR 2025 · similarity 0.88)
- Generative Region-Language Pretraining for Open-Ended Object Detection (CVPR 2024 · similarity 0.86)
- A Generative Approach at the Instance-Level for Image Segmentation Under Limited Training Data Conditions (Student Abstract) (AAAI 2025 · similarity 0.85)
- Semantic-Guided Generative Image Augmentation Method with Diffusion Models for Image Classification (AAAI 2024 · similarity 0.85)
- Compositional Scene Understanding through Inverse Generative Modeling (ICML 2025 · 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.
