Research map / Label-efficient and robust learning / Recognition and long-tail
End-to-End Deep Kronecker-Product Matching for Person Re-Identification
Yantao Shen, Tong Xiao, Hongsheng Li, Shuai Yi, Xiaogang Wang
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 KPM with soft feature warping (ablation) | Baseline ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 68.8 -> 70.6; 82.9 -> 85.5; 55.5 -> 56.0 | Improved |
| Remove soft warping after KPM (ablation) | Baseline+KPM · Market-1501, CUHK03, DukeMTMC | mAP: 68.8 -> 62.6; 82.9 -> 71.7; 55.5 -> 50.8 | Hurt |
| Add residual self-attention (ablation) | Baseline+KPM · Market-1501, CUHK03, DukeMTMC | mAP: 70.6 -> 72.0; 85.5 -> 86.2; 56.0 -> 58.1 | Improved |
| Add hourglass multi-scale features (ablation) | Baseline ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 68.8 -> 69.9; 82.9 -> 84.1; 55.5 -> 60.4 | Improved |
| Use separate scale losses and average scale confidences instead of concatenating scale features (ablation) | Baseline+HG · Market-1501 | mAP: 69.9 -> 69.1 | Hurt |
| Combine KPM, residual self-attention, and hourglass features (ablation) | Baseline ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 68.8 -> 75.3; 82.9 -> 89.2; 55.5 -> 63.2 | Improved |
| Use difference verification loss rather than cross-entropy, OIM, or triplet loss (ablation) | Base ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 59.8, 60.9, 67.9 -> 68.8; 62.7, 72.5, 80.7 -> 82.9; 40.7, 47.7, 54.6 -> 55.5 | Improved |
| Use 2:1 rather than 1:1 image aspect ratio with random initialization (ablation) | Base ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 30.2 -> 32.4; 60.8 -> 61.5; 23.2 -> 24.9 | Improved |
| Replace elementwise square with absolute value (ablation) | Base ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: +1.9%, +1.2%, +0.9% | Hurt |
| Move elementwise square before global average pooling (ablation) | Base ResNet-50 · Market-1501, CUHK03, DukeMTMC | mAP: 68.8 -> 64.3; 82.9 -> 72.7; 55.5 -> 50.9 | Hurt |
| Remove elementwise square (ablation) | Base ResNet-50 · Person re-identification | convergence | Hurt |
| Remove subtraction and elementwise square, and concatenate image-pair features (ablation) | Base ResNet-50 · Person re-identification | convergence | Hurt |
Closest papers by meaning
- Person Re-Identification With Deep Kronecker-Product Matching and Group-Shuffling Random Walk (TPAMI 2019 · similarity 0.90)
- Deep Neural Networks with Inexact Matching for Person Re-Identification (NeurIPS 2016 · similarity 0.89)
- Multi-Scale Deep Learning Architectures for Person Re-Identification (ICCV 2017 · similarity 0.88)
- Deeply-Learned Part-Aligned Representations for Person Re-Identification (ICCV 2017 · similarity 0.88)
- Interpretable and Generalizable Person Re-Identification with Query-Adaptive Convolution and Temporal Lifting (ECCV 2020 · similarity 0.87)
- Discriminative Feature Learning With Consistent Attention Regularization for Person Re-Identification (ICCV 2019 · similarity 0.87)
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.
