Research map / Label-efficient and robust learning / Knowledge distillation
Learning Compatible Embeddings
Qiang Meng, Chixiang Zhang, Xiaoqiang Xu, Feng Zhou
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 boundary loss to alignment loss (ablation) | LCE (alignment only, no transformation) · IJB-C (ResNet50, MS1Mv2) | upgrade gain (%): 60.82 -> 68.34 | Improved |
| add boundary loss to alignment loss (ablation) | LCE (alignment only, no transformation) · IJB-C (ResNet50, MS1Mv2) | performance gain (%): 84.28 -> 89.07 | Improved |
| add transformation module to alignment loss (ablation) | LCE (alignment only) · IJB-C (ResNet50, MS1Mv2) | performance gain (%): 84.28 -> 96.36 | Improved |
| add transformation module to alignment plus boundary loss (ablation) | LCE (alignment+boundary) · IJB-C (ResNet50, MS1Mv2) | performance gain (%): 89.07 -> 99.32 | Improved |
| add transformation module to alignment loss (ablation) | LCE (alignment only) · IJB-C (ResNet50, MS1Mv2) | upgrade gain (%): 60.82 -> 67.43 | Improved |
| add boundary loss to alignment plus transformation (ablation) | LCE (alignment+transformation) · IJB-C (ResNet50, MS1Mv2) | upgrade gain (%): 67.43 -> 72.44 | Improved |
| add transformation module to LCE (direct) (ablation) | LCE with ResNet18 new model · IJB-C (ResNet18, old ResNet50) | performance gain (%): -114.91 -> -95.63 | Mixed |
| use LCE point-to-set constraints instead of BCT | compatible training · IJB-C (training data change) | upgrade gain (%): 22.54 -> 81.37 | Improved |
| use LCE instead of BCT | compatible training · IJB-C (new backbone ResNet100) | upgrade gain (%): 52.04 -> 90.82 | Improved |
| use LCE instead of BCT | compatible training · IJB-C (new backbone MobileFace) | upgrade gain (%): -45.33 -> -47.50 | No clear effect |
| use LCE instead of BCT | compatible training · IJB-C (new loss CosFace) | upgrade gain (%): -75.38 -> -47.69 | Improved |
| use LCE instead of RBT when transformation trained with NormFace | cross model compatibility · IJB-C (phi*_a, phi*_b) | cross verification TAR@FAR=1e-4 (%): 92.78 -> 94.08 | Improved |
Closest papers by meaning
- Forward Compatible Training for Large-Scale Embedding Retrieval Systems (CVPR 2022 · similarity 0.90)
- Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements (CVPR 2024 · similarity 0.88)
- Compatibility-Aware Heterogeneous Visual Search (CVPR 2021 · similarity 0.88)
- Towards Backward-Compatible Representation Learning (CVPR 2020 · similarity 0.88)
- Continual Learning for Visual Search With Backward Consistent Feature Embedding (CVPR 2022 · similarity 0.87)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 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.
