Research map / Video / Video segmentation and tracking
Cross-video Identity Correlating for Person Re-identification Pre-training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CION, a cross-video identity-correlating pre-training framework for person ReID that denoises tracklets from internet videos into identities and pre-trains with identity-guided self-distillation, and releases the ReIDZoo model zoo of 32 models.
Problem
Person re-identification pre-training on person images from internet videos is limited to instance-level or single-video tracklet-level invariance, ignoring identity consistency of the same person across different videos.
Approach
Defines a noise concept combining intra-identity consistency and inter-identity discrimination, seeks cross-video identity correlation via progressive multi-level denoising (single-tracklet, short-range single-video, and long-range cross-video with a sliding range and linking relations), then applies an identity-guided DINO-style self-distillation loss contrasting augmented views of images from th
Key results
- With ResNet50-IBN and MGN fine-tuning, CION reaches 93.3% mAP on Market1501 and 74.3% mAP on MSMT17 using 3.9M pre-training images, surpassing ISR (47.8M images) by 1.0 and 2.8 mAP points.
- Average improvement across 32 models over ImageNet baselines under TransReID fine-tuning is 11.7% mAP on Market1501 and 18.8% mAP on MSMT17.
- On UDA MSMT17→Market1501 with ResNet50-IBN, CION reaches 92.1% mAP, versus 86.9% for LUP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CION pre-training instead of ISR pre-training | ResNet50-IBN · Market1501 supervised ReID | mAP: 1.0% | Helped |
| Use CION pre-training instead of ISR pre-training | ResNet50-IBN · MSMT17 supervised ReID | mAP: 2.8% | Helped |
| Add identity-level correlating to pre-training (ablation) | ResNet50-IBN with CION-AL and MGN · Market1501 | mAP: 91.4 -> 93.3; 90.9 -> 93.3 | Helped |
| Add identity-level correlating to pre-training (ablation) | ResNet50-IBN with CION-AL and MGN · MSMT17 | mAP: 68.2 -> 74.3; 66.7 -> 74.3 | Helped |
| Add single-tracklet denoising (ablation) | CION progressive multi-level denoising · Market1501 and MSMT17 | mAP | Helped |
| Add single-video denoising after single-tracklet denoising (ablation) | CION progressive multi-level denoising · Market1501 and MSMT17 | mAP | Helped |
| Add cross-video denoising after single-video denoising (ablation) | CION progressive multi-level denoising · Market1501 | mAP: +0.6% | Helped |
| Add cross-video denoising after single-video denoising (ablation) | CION progressive multi-level denoising · MSMT17 | mAP: +1.9% | Helped |
Surprising findings
- Pre-training with 3.9M images outperforms the ISR model pre-trained on 47.8M images, suggesting identity correlation matters more than data scale.
Limitations
- Fine-grained person representations are not explicitly learned.
- Automatically sought identity correlation is noisier than manual annotation due to feature extractor limits.
- Sliding range may miss correlations between tracklets of the same person that are far apart in a video.
Experimental setup
Datasets: SYNTH-PEDES LUPerson UFine6926 Market1501 MSMT17
Benchmarks: Market1501 MSMT17
Metrics: mAP Rank-1
Baselines: LUP LUP-NL PASS ISR DINO TranSSL SOLIDER C-Contrast
Builds on: DINO LUP TransReID ISR PLIP
Closest papers by meaning
- Large-Scale Pre-Training for Person Re-Identification With Noisy Labels (CVPR 2022 · similarity 0.86)
- Asymmetric Co-Teaching for Unsupervised Cross-Domain Person Re-Identification (AAAI 2020 · similarity 0.86)
- AXM-Net: Implicit Cross-Modal Feature Alignment for Person Re-identification (AAAI 2022 · similarity 0.85)
- From Poses to Identity: Training-Free Person Re-Identification via Feature Centralization (CVPR 2025 · similarity 0.84)
- Diffusion-based Synthetic Data Generation for Visible-Infrared Person Re-Identification (AAAI 2025 · similarity 0.83)
- Temporal Correlation Vision Transformer for Video Person Re-Identification (AAAI 2024 · similarity 0.83)
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.