Research map / Label-efficient and robust learning / Recognition and long-tail
Attribute-Guided Pedestrian Retrieval: Bridging Person Re-ID with Internal Attribute Variability
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.
Introduces the Attribute-Guided Pedestrian Retrieval (AGPR) task and the ATPR transformer, which fuses global ID learning with attribute-grouped local branches and correlation/decorrelation regularization, evaluated on a new RAP-based benchmark.
Problem
Standard person re-identification ignores intra-ID attribute variations (e.g., changed clothing or accessories), so retrieval cannot be refined by user-specified attributes while keeping identity matching.
Approach
Combines a ViT-based global ID branch with a local attribute branch where tokens are grouped by attribute groups via OpenPose keypoints and attribute embeddings, plus inter-group correlation and intra-group decorrelation regularizers, so queries can be retrieved with specified attributes while preserving identity.
Key results
- ATPR reaches 74.26% mAP and 86.93% rank-1 on RAP without attribute specification, outperforming TransReID+CE+JPM (71.39% / 83.77%).
- Under 90% cross-attribute retrieval, ATPR attains 49.51% mAP and 60.33% rank-1, versus 38.91% mAP for TransReID.
- Ablation: removing the attribute embedding (w/o adding AE) drops mAP to 68.85% without cross attributes, showing attribute embedding matters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ATPR rather than TransReID+CE+JPM | RAP person retrieval · RAP, no cross att. | mAP: 71.39 -> 74.26 | Helped |
| Use ATPR rather than TransReID+CE+JPM | RAP person retrieval · RAP, no cross att. | rank-1: 83.77 -> 86.93 | Helped |
| Use ATPR rather than TransReID | RAP retrieval with specified attributes · RAP, cross-att.:90% | mAP: 38.91 -> 49.51 | Helped |
| Use ATPR rather than TransReID | RAP retrieval with specified attributes · RAP, cross-att.:90% | rank-1: 42.76 -> 60.33 | Helped |
| Remove the global loss, leaving only local loss (ablation) | ATPR · RAP, no cross att. | mAP: 74.26 -> 69.98 | Hurt |
| Remove the local loss, leaving only global loss (ablation) | ATPR · RAP, no cross att. | mAP: 74.26 -> 70.47 | Hurt |
| Remove attribute regularization (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 46.66 | Hurt |
| Use IGAC without IGAD (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 48.41 | Hurt |
| Use IGAD without IGAC (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 48.67 | Hurt |
| Use fixed patch-attribute associations (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 48.46 | Hurt |
| Remove attribute embeddings (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 43.87 | Hurt |
| Remove the [class] token from AEG (ablation) | ATPR · RAP, cross-att.:90% | mAP: 49.51 -> 45.22 | Hurt |
Surprising findings
- Clothing-change Re-ID methods (ReIDCaps, CAL) underperform the traditional HACNN baseline on RAP.
Limitations
- Potential weakness in scenarios with unseen attributes or significant occlusions, as noted by the authors.
- Benchmark built on a single dataset (RAP) with manually grouped attributes.
Experimental setup
Datasets: RAP
Benchmarks: RAP AGPR benchmark
Metrics: mAP rank-1
Baselines: HACNN PAR ReIDCaps CAL TransReID IDE-ATT-R MSCAN
Builds on: TransReID ViT CAL
Closest papers by meaning
- Learning Heterogeneous Dictionary Pair with Feature Projection Matrix for Pedestrian Video Retrieval via Single Query Image (AAAI 2017 · similarity 0.87)
- Selective and Orthogonal Feature Activation for Pedestrian Attribute Recognition (AAAI 2024 · similarity 0.86)
- FlexiReID: Adaptive Mixture of Expert for Multi-Modal Person Re-Identification (ICML 2025 · similarity 0.85)
- Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework (AAAI 2025 · similarity 0.85)
- Cheb-GR: Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification (CVPR 2025 · similarity 0.84)
- Enhanced Visual-Semantic Interaction with Tailored Prompts for Pedestrian Attribute Recognition (CVPR 2025 · similarity 0.84)
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.