CA-ReID

CVPR 2026

Composite-Attribute Person Re-Identification via Pose-Guided Disentanglement

Kartik Patwari1* Noranart Vesdapunt2 Chien-Yi Wang2 Dawei Li2 Cong Phuoc Huynh2 Ning Zhou2 Chen-Nee Chuah1 Kah Kuen Fu2

1University of California, Davis 2Amazon

*This work was done at Amazon.

Paper BibTeX
CA-ReID retrieval comparison
Example retrievals under a short attribute query.

Problem. Existing multimodal Re-ID methods perform well with full descriptions, but their accuracy drops substantially for short attribute queries. CA-ReID studies retrieval conditioned on a reference image and a short or composite attribute query, while requiring both identity consistency and attribute satisfaction.

A New Setting

Composed Attribute Person ReID (CA-ReID)

The query consists of a reference person image and a short attribute condition.

CA-ReID query setting and ambiguity levels
CA-ReID organizes query conditions into easy, medium, and hard levels based on how many attributes are specified.

Method

CA-ReID Method Overview

CA-ReID framework figure
The pipeline builds part-aware image and text slots, fuses them into a composed representation, and retrieves the ranked gallery from that query.

Part-Aware Representations (PAR)

Pose groups image patches into body regions, and text is projected into matching slots.

Dense Disentangling Loss (DDL)

DDL separates identity cues from attribute edits and reduces cross-part leakage.

Composed Retrieval

The reference image and attribute query are fused, then matched against the gallery.

Benchmark

CA-ReID Benchmark

CA-ReID dataset attribute examples
Benchmark attributes cover head, top, bottom, feet, belongings, and context cues.

Results

Quantitative Results

CA-ReID improves retrieval accuracy on short-query settings, with the largest gains on the harder splits.

CA-ReID benchmark summary

Results on the proposed CA-ReID setting.

Method Query Celeb-ReID-L COCAS+Real2
R@1 R@5 mAP R@1 R@5 mAP
DIFFER [38]* E 23.6 25.2 12.5 31.6 39.2 14.9
Inst-ReID [21] E 81.8 95.6 20.8 82.7 93.5 36.4
M 74.0 81.8 19.0 51.8 78.2 17.9
H 41.6 71.0 14.7 44.0 67.9 19.9
CA-ReID (Ours) E 83.1 97.8 24.5 83.9 94.7 38.3
M 78.9 86.2 23.3 55.1 82.4 21.2
H 58.6 79.4 20.4 50.4 74.2 21.3

*Image-only ReID.

Hard queries by attribute region

Celeb-ReID-L results for CA-ReID (Ours).

Region R@1 R@5 mAP
Head 59.4 79.3 18.4
Top 62.4 85.2 23.5
Bottom 60.3 83.2 21.9
Feet 58.3 80.2 21.0
Accessories 55.4 76.1 19.5
Belongings 58.2 77.8 19.9
Context 56.2 74.1 18.4

Standard CC-ReID benchmark comparisons

Comparison on LTCC and PRCC benchmarks.

Method Venue LTCC PRCC
Top1 mAP Top1 mAP
TransReID CVPR'21 46.6 44.8 34.4 17.1
CAL CVPR'22 55.2 55.8 40.1 18.0
AIM CVPR'23 57.9 58.3 40.6 19.1
LDF ACM'23 58.4 58.6 32.9 15.4
3DInv ICCV'23 40.9 18.9 56.5 57.2
CCFA CVPR'23 45.3 22.1 61.2 58.4
CLIP3D CVPR'24 42.1 22.9 61.8 58.3
Inst-ReID CVPR'24 66.7 46.7 54.2 52.3
DIFFER CVPR'25 68.5 64.7 58.2 31.6
CA-ReID (Ours) CVPR'26 63.8 53.7 55.2 43.4

Examples

Qualitative Retrieval Examples

Short-query retrieval examples from the benchmark.

CA-ReID retrieval examples
Retrieval examples showing where the target attribute is satisfied and where identity-only matches fail.

BibTeX

Citation

@InProceedings{patwari2026careid,
  title     = {Composite-Attribute Person Re-Identification via Pose-Guided Disentanglement},
  author    = {Patwari, Kartik and Vesdapunt, Noranart and Wang, Chien-Yi and Li, Dawei and
               Huynh, Cong Phuoc and Zhou, Ning and Chuah, Chen-Nee and Fu, Kah Kuen},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026}
}