Research map / Label-efficient and robust learning / Recognition and long-tail
Multi-Centroid Representation Network for Domain Adaptive Person Re-ID
Yuhang Wu, Tengteng Huang, Haotian Yao, Chi Zhang, Yuanjie Shao, Chuchu Han, Changxin Gao, Nong Sang
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 uni-centroid cluster representation with multi-centroid memory (ablation) | SpCL baseline · Duke→Market, Market→Duke, Duke→MSMT, Market→MSMT | mAP: 73.1 -> 81.5; 62.1 -> 70.8; 19.1 -> 30.2; 18.2 -> 28.1 | Improved |
| add DSCL, restricting contrastive comparisons to the query's domain (ablation) | MCM · Duke→Market, Market→Duke, Duke→MSMT, Market→MSMT | mAP: 81.5 -> 82.4; 70.8 -> 71.4; 30.2 -> 31.1; 28.1 -> 29.7 | Improved |
| add SONI synthetic hard negatives (ablation) | MCM · Duke→Market, Market→Duke, Duke→MSMT, Market→MSMT | mAP: 81.5 -> 82.9; 70.8 -> 70.8; 30.2 -> 33.7; 28.1 -> 32.3 | Improved |
| choose the most-similar positive centroid rather than the moderate one (ablation) | MCM · Duke→Market | mAP: 81.5 -> 56.3 | Hurt |
| choose the least-similar positive centroid rather than the moderate one (ablation) | MCM · Duke→Market, Market→Duke | mAP: 81.5 -> 79.9; 70.8 -> 69.4 | Hurt |
| use all negative centroids instead of their per-cluster mean (ablation) | MCM with moderate positive selection · Duke→Market, Market→Duke | mAP: 81.5 -> 81.0; 70.8 -> 70.0 | Hurt |
| use mean negative centroids rather than all negative centroids (ablation) | MCM with moderate positive selection · Duke→Market, Market→Duke | mAP: 81.0 -> 81.5; 70.0 -> 70.8 | Improved |
| increase centroid count per class from 2 to 4 (ablation) | MCRN · Duke→Market | mAP: 75.7 -> 83.8 | Improved |
| increase centroid count per class above 4 (ablation) | MCRN · Duke→Market | mAP: 83.8; 83.5; 83.5 | No clear effect |
| use QNNI interpolation between the query and a nearest negative instead of SONI (ablation) | MCRN · Duke→Market, Market→Duke | mAP: 83.8 vs 76.5; 71.5 vs 66.1 | Hurt |
| use RNNI interpolation between two randomly selected top-nearest negatives instead of SONI (ablation) | MCRN · Duke→Market, Market→Duke | mAP: 83.8 vs 83.4; 71.5 vs 70.4 | Hurt |
| add MCM on source-domain data (ablation) | baseline · Duke→Market | mAP: +1.8% | Improved |
Closest papers by meaning
- Towards Discriminative Representation Learning for Unsupervised Person Re-Identification (ICCV 2021 · similarity 0.90)
- Unsupervised Multi-Source Domain Adaptation for Person Re-Identification (CVPR 2021 · similarity 0.90)
- IDM: An Intermediate Domain Module for Domain Adaptive Person Re-ID (ICCV 2021 · similarity 0.89)
- Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification (ICLR 2020 · similarity 0.88)
- Domain Adaptive Attention Learning for Unsupervised Person Re-Identification (AAAI 2020 · similarity 0.88)
- Unsupervised Domain Adaptation with Noise Resistible Mutual-Training for Person Re-identification (ECCV 2020 · 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.
