Research map / Label-efficient and robust learning / Recognition and long-tail
Omni-Scale Feature Learning for Person Re-Identification
Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, Tao Xiang
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 unified aggregation gate instead of plain addition (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 92.0 -> 93.6 | Improved |
| add unified aggregation gate instead of concatenation (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 91.4 -> 93.6 | Improved |
| increase stream cardinality from T=1 to T=4 (ablation) | OSNet · Market1501 (ablation) | Rank-1 accuracy: 86.5 -> 93.6 | Improved |
| use streams of different depth (omni-scale) instead of same depth (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 91.7 -> 93.6 | Improved |
| share one gate across streams instead of separate gates (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 92.9 -> 93.6 | Improved |
| use channel-wise gates instead of stream-wise gates (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 92.6 -> 93.6 | Improved |
| use dynamic input-dependent gates instead of static learned-and-fixed gates (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 91.6 -> 93.6 | Improved |
| use factorised Lite 3x3 convolutions instead of full convolutions (ablation) | OSNet T=4 · Market1501 (ablation) | Rank-1 accuracy: 94.0 -> 93.6 | No clear effect |
| add unified AG with four streams instead of single scale (ablation) | OSNet · CIFAR10 | error rate (%): 5.49 -> 4.41 | Improved |
| add deep mutual learning (model-1) (ablation) | OSNet · Market1501 (training recipe, cosine annealing) | mAP: 86.5 -> 87.2 | Improved |
| add triplet loss with weight 0.5 (ablation) | OSNet · Market1501 (training recipe, cosine annealing) | Rank-1 accuracy: 94.6 -> 95.5 | Improved |
| add entropy maximisation regulariser (lambda 0.1) (ablation) | OSNet · Market1501 (training recipe, cosine annealing) | Rank-1 accuracy: 94.6 -> 94.7 | No clear effect |
Closest papers by meaning
- Learning Generalisable Omni-Scale Representations for Person Re-Identification (TPAMI 2021 · similarity 0.97)
- Multi-Scale Deep Learning Architectures for Person Re-Identification (ICCV 2017 · similarity 0.91)
- Leader-Based Multi-Scale Attention Deep Architecture for Person Re-Identification (TPAMI 2019 · similarity 0.91)
- Deep Learning for Person Re-Identification: A Survey and Outlook (TPAMI 2021 · similarity 0.90)
- Interaction-And-Aggregation Network for Person Re-Identification (CVPR 2019 · similarity 0.89)
- Discriminative Feature Learning With Consistent Attention Regularization for Person Re-Identification (ICCV 2019 · similarity 0.89)
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.
