Research map / Video / Video segmentation and tracking
CLIMB-ReID: A Hybrid CLIP-Mamba Framework for Person Re-Identification
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 CLIMB-ReID, a CLIP-plus-Mamba hybrid for image and video person re-identification that uses parameter-free, prompt-free multi-memory collaboration to transfer CLIP knowledge and importance-aware reorder Mamba for multi-temporal spatiotemporal modeling.
Problem
Person re-identification with CLIP suffers from costly transfer of CLIP knowledge (prompt learning adds parameters) and inefficient capture of spatiotemporal context in videos (self-attention has quadratic cost).
Approach
Replace text prompts with multiple identity-specific visual memories updated from training batches (Mean/Hard selection) with a multi-memory collaboration loss, and model multi-stride frame sequences with Mamba after reordering patch tokens by importance (cosine similarity to the class token).
Key results
- Achieves 89.7% mAP and 93.3% Rank-1 on MARS, surpassing TF-CLIP by 0.3 mAP and 0.3 Rank-1.
- Reaches 92.6% mAP and 96.8% Rank-1 on Market-1501 and 77.8% mAP and 90.5% Rank-1 on MSMT17, outperforming PCL and TF-CLIP.
- On LS-VID, reaches 85.0% mAP and 91.3% Rank-1, with the full model improving over the TF-CLIP-style baseline by 4.7 mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CLIMB-ReID instead of TF-CLIP | video person ReID · MARS | mAP: 89.4 -> 89.7 | Helped |
| use CLIMB-ReID instead of TF-CLIP | video person ReID · LS-VID | Rank-1: 90.4 -> 91.3 | Helped |
| use CLIMB-ReID instead of TF-CLIP | video person ReID · iLIDS-VID | Rank-1: 94.5 -> 96.7 | Helped |
| use CLIMB-ReID instead of PCL | image person ReID · Market1501 | mAP: 91.4 -> 92.6 | Helped |
| use CLIMB-ReID instead of PCL | image person ReID · MSMT17 | Rank-1: 89.8 -> 90.5 | Helped |
| add MMC to the CLIP-Memory baseline (ablation) | video person ReID · LS-VID | mAP: 80.3 -> 82.0 | Helped |
| add MMC to the CLIP-Memory baseline (ablation) | image person ReID · MSMT17 | mAP: 73.9 -> 76.7 | Helped |
| add IRM at slice stride 1 to MMC (ablation) | CLIMB-ReID · LS-VID | mAP: 82.0 -> 83.2 | Helped |
| add MTM with temporal strides larger than 1 to MMC (ablation) | CLIMB-ReID · LS-VID | mAP: 82.0 -> 84.4 | Helped |
| add IRM to MMC + MTM (ablation) | CLIMB-ReID · LS-VID | Rank-1: 91.0 -> 91.3 | Helped |
| use Mean+Hard memory updates instead of Mean alone (ablation) | MMC · LS-VID | mAP: 79.2 -> 82.0 | Helped |
| use Mean+Hard memory updates instead of Mean alone (ablation) | MMC · MARS | mAP: 88.4 -> 88.6 | Helped |
Surprising findings
- Pure Mamba structures (VMamba, VideoMamba) perform far below expectations on MARS, suggesting person ReID may need specialized pre-training.
Limitations
- Pure Mamba baselines underperform, implying dependence on CLIP backbone and additional pre-training.
- Efficiency claims are not quantified with measured compute or runtime in the main text.
Experimental setup
Datasets: MARS LS-VID iLIDS-VID Market-1501 MSMT17
Benchmarks: MARS LS-VID iLIDS-VID Market-1501 MSMT17
Metrics: mAP Rank-1 Rank-5 CMC
Baselines: CLIP-ReID TF-CLIP PCL TMT TCVIT CAVIT TransReID VMamba VideoMamba
Builds on: CLIP Mamba TF-CLIP CLIP-ReID VideoMamba
Closest papers by meaning
- TF-CLIP: Learning Text-Free CLIP for Video-Based Person Re-identification (AAAI 2024 · similarity 0.90)
- Distilling CLIP with Dual Guidance for Learning Discriminative Human Body Shape Representation (CVPR 2024 · similarity 0.87)
- MambaPro: Multi-Modal Object Re-identification with Mamba Aggregation and Synergistic Prompt (AAAI 2025 · similarity 0.87)
- Exploring the Camera Bias of Person Re-identification (ICLR 2025 · similarity 0.87)
- Instruct-ReID: A Multi-purpose Person Re-identification Task with Instructions (CVPR 2024 · similarity 0.86)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.86)
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.