Research map / Architectures and efficiency / Pruning and efficient tuning
Backbone Cannot Be Trained at Once: Rolling Back to Pre-Trained Network for Person Re-Identification
Read the paperCheck 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 |
|---|---|---|---|
| Apply rolling-back refine-tuning instead of ordinary fine-tuning | ResNet-50 · Market-1501 | mAP: 73.16 -> 77.0 | Helped |
| Apply rolling-back refine-tuning instead of ordinary fine-tuning | ResNet-50 · DukeMTMC | mAP: 63.26 -> 66.6 | Helped |
| Apply rolling-back refine-tuning instead of ordinary fine-tuning | ResNet-50 · CUHK03-L | mAP: 45.17 -> 50.7 | Helped |
| Apply rolling-back refine-tuning instead of ordinary fine-tuning | ResNet-50 · CUHK03-D | mAP: 44.05 -> 47.4 | Helped |
| Add rolling-back refinement periods that progressively keep low-level blocks tuned (ablation) | ResNet-50 · Market-1501 | mAP: 73.16 -> 75.65 -> 76.54 -> 77.01 | Helped |
| Add the fourth refinement period, keeping B1, B2, B3, and FC tuned (ablation) | ResNet-50 · DukeMTMC | mAP: 66.57 -> 66.39 | Hurt |
| Retain Block1 while rolling back the other backbone blocks (ablation) | ResNet-50 · Market-1501 | mAP: 73.16 -> 74.08 | Helped |
| Retain Block2 while rolling back the other backbone blocks (ablation) | ResNet-50 · Market-1501 | mAP: 73.16 -> 74.37 | Helped |
| Retain Block5 while rolling back the other backbone blocks (ablation) | ResNet-50 · Market-1501 | mAP: 73.16 -> 71.17 | Hurt |
| Use rolling back rather than learning-rate-only perturbation (ablation) | ResNet-50 fine-tuning · Market-1501 and DukeMTMC | mAP | Helped |
| Fine-tune the backbone during the initial period instead of freezing it for FC warm-up (ablation) | ResNet-50 fine-tuning · Market-1501 and DukeMTMC | mAP | Helped |
| Use rolling-back refinement with ResNet-34 or ResNet-101 instead of a single fine-tuning period (ablation) | ResNet-34 and ResNet-101 · Market-1501 and DukeMTMC | mAP | Helped |
Closest papers by meaning
- A Multi-Task Deep Network for Person Re-Identification (AAAI 2017 · similarity 0.92)
- Unleashing Potential of Unsupervised Pre-Training With Intra-Identity Regularization for Person Re-Identification (CVPR 2022 · similarity 0.89)
- Resource Aware Person Re-Identification Across Multiple Resolutions (CVPR 2018 · similarity 0.89)
- Efficient and Deep Person Re-Identification Using Multi-Level Similarity (CVPR 2018 · similarity 0.89)
- Interaction-And-Aggregation Network for Person Re-Identification (CVPR 2019 · similarity 0.89)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · 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.