Research map / Label-efficient and robust learning / Knowledge distillation
Distilled Person Re-Identification: Towards a More Scalable System
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| distill teacher pairwise similarities with the proposed Log-Euclidean loss rather than Hinton et al.'s soft-label loss | MobileNetV2 Re-ID student · Market-1501 and DukeMTMC | rank-1 accuracy (%): 41.2 -> 49.7; 32.5 -> 47.6 | Helped |
| use the Log-Euclidean similarity loss rather than PKT | MobileNetV2 Re-ID student · Market-1501 and DukeMTMC | mAP (%): 22.3 -> 24.6; 28.6 -> 31.1 | Helped |
| add the matrix logarithm to the similarity distillation loss (ablation) | Log-Euclidean similarity distillation · Market-1501 | rank-1 accuracy (%): 47.7 -> 49.7 | Helped |
| add the matrix logarithm to the similarity distillation loss (ablation) | Log-Euclidean similarity distillation · DukeMTMC | mAP (%): 29.8 -> 31.1 | Helped |
| use adaptive teacher weights learned from a few labelled validation identities instead of fixed equal weights (ablation) | multi-teacher similarity distillation · DukeMTMC | rank-1 accuracy (%): 48.4 -> 57.4 | Helped |
| increase validation identities from 0 to 1 for learning teacher weights (ablation) | adaptive knowledge aggregator · DukeMTMC | rank-1 accuracy (%): 48.4 -> 55.9 | Helped |
| increase validation identities from 1 to 5 for learning teacher weights (ablation) | adaptive knowledge aggregator · Market-1501 | rank-1 accuracy (%): 62.6 -> 63.2 | Helped |
| use a ResNet-18 rather than MobileNetV2 student backbone (ablation) | proposed Re-ID method · DukeMTMC | mAP (%): 36.7 -> 36.4 | Hurt |
| use a ResNet-50 rather than MobileNetV2 student backbone (ablation) | proposed Re-ID method · DukeMTMC | rank-1 accuracy (%): 57.4 -> 56.1 | Hurt |
| initialize with the proposed method rather than ImageNet pretraining before fine-tuning on 20% of identities (ablation) | MobileNetV2 Re-ID student · Market-1501 | rank-1 accuracy (%): 60.1 -> 77.1 | Helped |
| initialize with the proposed method rather than ImageNet pretraining before fine-tuning on 20% of identities (ablation) | MobileNetV2 Re-ID student · DukeMTMC | mAP (%): 29.7 -> 46.0 | Helped |
| distill from multiple selected teachers rather than only the top-weight teacher (ablation) | multi-teacher student for Market-1501 · Market-1501 | rank-1 accuracy (%): 48.1 -> 63.0 | Helped |
Closest papers by meaning
- Unsupervised Domain Adaptation for Person Re-identification via Heterogeneous Graph Alignment (AAAI 2021 · similarity 0.89)
- Meta Batch-Instance Normalization for Generalizable Person Re-Identification (CVPR 2021 · similarity 0.89)
- Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification (ICLR 2020 · similarity 0.89)
- Camera-Aware Proxies for Unsupervised Person Re-Identification (AAAI 2021 · similarity 0.89)
- Id-Free Person Similarity Learning (CVPR 2022 · similarity 0.89)
- Intra-Inter Camera Similarity for Unsupervised Person Re-Identification (CVPR 2021 · similarity 0.88)
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.