Research map / Vision-language and multimodal / Cross-modal learning
MambaPro: Multi-Modal Object Re-identification with Mamba Aggregation and Synergistic Prompt
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 MambaPro, which adapts a frozen CLIP to multi-modal (RGB/NIR/TIR) object ReID using a parallel feed-forward adapter, synergistic residual prompts, and a Mamba-based aggregation module with linear complexity.
Problem
Multi-modal object re-identification with RGB, NIR and TIR images, where CLIP is unexplored and attention-based multi-modal aggregation has quadratic cost on long token sequences.
Approach
Combine a parallel adapter on CLIP's FFN, modality-synergistic residual prompts that transfer prompts across modalities layer by layer, and a Mamba (selective SSM) aggregation that models intra- and inter-modality tokens with linear complexity.
Key results
- On RGBNT201, MambaPro reaches 78.9% mAP and 83.4% Rank-1, surpassing TOP-ReID (72.3% mAP) by 6.6 points.
- On MSVR310, MambaPro reaches 47.0% mAP, surpassing EDITOR (39.0%) by 8.0 points.
- On RGBNT100, MambaPro reaches 83.9% mAP, surpassing EDITOR (82.1%) by 1.8 points, with 74.20M trainable parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MambaPro instead of TOP-ReID | Multi-modal object ReID · RGBNT201 | mAP: 72.3 -> 78.9 | Helped |
| Use MambaPro instead of RSCNet | Multi-modal object ReID · RGBNT100 | mAP: 82.3 -> 83.9 | Helped |
| Use MambaPro instead of RSCNet | Multi-modal object ReID · MSVR310 | mAP: 39.5 -> 47.0 | Helped |
| Add PFA to frozen CLIP (ablation) | MambaPro · RGBNT201 | mAP: 69.4 -> 73.2 | Helped |
| Add SRP to frozen CLIP (ablation) | MambaPro · RGBNT201 | mAP: 69.4 -> 71.3 | Helped |
| Add MA to frozen CLIP with PFA and SRP (ablation) | MambaPro · RGBNT201 | mAP: 74.4 -> 78.9 | Helped |
| Use synergistic prompt transfer instead of independent prompts (ablation) | CLIP(L) prompt tuning · RGBNT201 | mAP: 64.7 -> 68.9 | Helped |
| Add residual prompt connections to SP (ablation) | CLIP(L) prompt tuning · RGBNT201 | mAP: 68.9 -> 70.1 | Helped |
| Fuse residual prompts by averaging modality tokens (ablation) | SRP · RGBNT201 | mAP: 70.1 -> 71.3 | Helped |
| Replace PFA with a bottleneck adapter (ablation) | CLIP(L) · RGBNT201 | mAP: 68.5 -> 73.2 | Hurt |
| Use both intra-modality and inter-modality MA instead of either alone (ablation) | MambaPro · RGBNT201 | mAP: 77.5 -> 78.9 | Helped |
| Insert PFA beside FFN rather than behind FFN (ablation) | CLIP(L) + PFA · RGBNT201 | mAP: 46.9 -> 73.2 | Helped |
Surprising findings
- Fully fine-tuning CLIP (CLIP(F)) gives comparable results (78.7% mAP) to the frozen-backbone variant (78.9% mAP) on RGBNT201.
- Sequential placement of the adapter after MHSA or behind FFN performs much worse (e.g., 46.9% mAP for behind FFN) than the parallel side-of-FFN placement (73.2%).
Limitations
- Results on RGBNT100 Rank-1 (94.7%) trail some baselines such as LRMM (96.7%).
- Hyperparameter and module choices are tuned on RGBNT201 and the vehicle generalization analysis is limited to supplementary material.
Experimental setup
Datasets: RGBNT201 RGBNT100 MSVR310
Benchmarks: RGBNT201 RGBNT100 MSVR310
Metrics: mAP CMC Rank-1 Rank-5 Rank-10 FLOPs parameters GPU memory
Baselines: TOP-ReID EDITOR RSCNet HTT TransReID LRMM UniCat IEEE
Builds on: CLIP Mamba Vision Mamba VMamba TOP-ReID
Closest papers by meaning
- Magic Tokens: Select Diverse Tokens for Multi-modal Object Re-Identification (CVPR 2024 · similarity 0.88)
- Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.88)
- DeMo: Decoupled Feature-Based Mixture of Experts for Multi-Modal Object Re-Identification (AAAI 2025 · similarity 0.88)
- CLIMB-ReID: A Hybrid CLIP-Mamba Framework for Person Re-Identification (AAAI 2025 · similarity 0.87)
- IDEA: Inverted Text with Cooperative Deformable Aggregation for Multi-modal Object Re-Identification (CVPR 2025 · similarity 0.86)
- MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification (NeurIPS 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.