In PRAISE of Attributes: Grounding Retrieval Geometry with Query-Conditioned Subspaces for Person Re-Identification

Rajarshi Bhattacharya; Aryan Shukla; Shakeeb Murtaza; Christian Desrosiers; Jose Dolz; Maguelonne Heritier; Eric Granger

cs.CVOct 20, 2026

In PRAISE of Attributes: Grounding Retrieval Geometry with Query-Conditioned Subspaces for Person Re-Identification

Authors: Rajarshi Bhattacharya, Aryan Shukla, Shakeeb Murtaza, Christian Desrosiers, Jose Dolz, Maguelonne Heritier, Eric Granger

Organizations: LIVIA, École de technologie supérieure · Genetec Inc.

Abstract

Person re-identification (ReID) ranks an image gallery using either an image query or a textual query. In image-to-image (I2I) ReID, viewpoint, illumination, occlusion, and background changes can make distractors appear more similar than the true match. In text-to-image (T2I) ReID, incomplete or ambiguous descriptions may omit key visual evidence. We study this problem for vision-language model (VLM) based ReID, where attributes such as clothing color and carried objects provide useful fine-grained matching cues. However, directly using attribute-prompt similarities as retrieval scores is unreliable because templated prompts can be biased, correlated, and poorly calibrated. We therefore propose PRompt-guided Attribute-based Inference-time Similarity Enhancement (PRAISE), a query-side test-time adaptation framework for VLM-based ReID. Instead of treating prompts as scores, PRAISE uses a transferable attribute-prompt bank to construct an Attribute-Conditioned Subspace (ACS): a query-dependent low-rank subspace induced by the most compatible attribute prompts. ACS provides a controlled attribute-conditioned geometry for reshaping query--gallery affinities and guiding lightweight inference-time updates without identity or attribute labels. Experiments across multiple I2I- and T2I-ReID benchmarks show that PRAISE consistently improves VLM retrieval backbones and outperforms both generic and ReID-specific TTA baselines.

Figures & tables

Figure 1: Motivation and overview of PRAISE . (a) In I2I-ReID, raw visual similarity can retrieve distractors caused by camera, pose, or background bias. (b) In T2I-ReID, incomplete text descriptions can produce partial matches that share only some mentioned attributes. (c) PRAISE addresses both cases by selecting query-relevant attribute directions, orthogonalizing them into an Attribute-Conditioned Subspace (ACS), and using this controlled subspace to reshape query-gallery affinities during test-time.

Panel of Figure 2 — Figure 2: Similarity distributions for matching and non-matching query-gallery pairs on Market-1501. Direct prompt-profile similarity (a) is saturated for both groups, whereas ACS (b) yields clearer separation between matches and distractors.

Figure 3: Method overview. PRAISE uses attribute prompts to construct a query-specific semantic subspace that stabilizes query-side adaptation under domain shift. At test time, it performs query-side episodic TTA with a fixed source-space gallery index. Gallery and prompt-bank embeddings are computed once and kept fixed; gradients update only the query-side normalization affine parameters.

Backbone	I2IReID				T2IReID
	MSMT17		Market-1501		CUHK-Pedes		RSTPReID
	mAP	R1	mAP	R1	mAP	R1	mAP	R1
CLIP ViT-B/16 [ 35 ]	0 1.1 $/$ 0 1.4	0 5.6 $/$ 0 6.7	0 3.5 $/$ 0 4.3	11.9 $/$ 14.0	0 7.1 $/$ 0 9.1	0 8.9 $/$ 11.7	0 8.1 $/$ 0 9.2	11.2 $/$ 12.9
DINO.txt [ 14 ]	0 0.7 $/$ 0 1.8	0 5.2 $/$ 0 9.4	0 1.9 $/$ 0 3.0	0 7.1 $/$ 10.1	0 7.3 $/$ 0 8.8	0 9.1 $/$ 12.8	10.3 $/$ 13.1	11.9 $/$ 14.6
CLIPReID [ 26 ]	20.8 $/$ 24.7	45.9 $/$ 52.8	37.8 $/$ 42.1	63.3 $/$ 69.4	0 8.1 $/$ 0 9.9	0 8.2 $/$ 11.4	12.2 $/$ 13.8	13.5 $/$ 15.1
RDE [ qin2024noisy ]	-	-	-	-	33.5 $/$ 38.1	36.8 $/$ 42.8	41.2 $/$ 46.3	54.9 $/$ 60.9

Table 1: Effect on different backbones. Backbones evaluated on I2IReID and T2IReID datasets with PRAISE applied (score on left side of ‘

/

’ is without adaptation and score on right is after applying PRAISE ). Each dataset reports mAP and Rank-1 Accuracy (R1).

TTA Method	Type	Dataset				Time (ms) / sample
		MSMT17		Market-1501
		mAP	R1	mAP	R1
No-adapt (Source) [NeurIPS’24]	–	10.5	34.2	35.0	62.6	0 1.8
TENT [ wang2020tent ] [ICLR’21]	Classifier logit	0 9.2	29.8	29.3	52.7	16.2
SHOT [ 27 ] [ICML’20]		0 9.9	30.2	28.9	56.8	63.8
SAR [ 29 ] [ICLR’23]		10.0	31.1	33.2	63.3	91.4

Table 2: Comparison of TTA methods on I2IReID benchmarks using a PLIP [ 55 ] backbone. Each dataset reports mAP and Rank-1 Accuracy (R1).

TTA Method	Type	Dataset				Time (ms) / sample
		CUHK-Pedes		RSTPReID
		mAP	R1	mAP	R1
No-adapt (Source) [NeurIPS’24]	–	46.4	55.9	23.1	34.5	12.9
TENT [ wang2020tent ] [ICLR’21]	Classifier logit	46.4	56.2	23.3	34.2	81.8
SHOT [ 27 ] [ICML’20]		47.1	56.6	24.2	34.0	179.1
TPT [ 36 ] [NeurIPS’22]	VLM based	46.5	55.9	23.1	34.6	4193.4

Table 3: Comparison of TTA methods on T2IReID benchmarks on a PLIP [ 55 ] backbone. Each dataset reports mAP and Rank-1 Accuracy (R1).

Panel of Figure 4 — Figure 4: Hyperparameter Sensitivity. Performance with respect to varying hyperparameters. Tested on Market-1501.

Group	Variant	I2I-ReID		T2I-ReID
Group	Variant	mAP	R1	mAP	R1
Component Ablation	No adapt	35.0	62.6	23.1	34.5
	ACS only (No-adaptation, $\alpha_{\mathrm{acs}}{=}0.1$ )	35.2	61.8	23.6	34.1
	Adaptation w/o ACS $(\alpha_{\mathrm{acs}}{=}0)$	36.8	65.0	23.5	35.4
	ACS adaptation (i.e. PRAISE )	46.5	71.1	27.6	37.6
Subspace Ablation	Shuffled prompts across queries	37.8	62.8	23.5	35.3

Table 4: Ablations. All rows use the same TopR, rank

r

, number of adaptation steps, and

\alpha_{\mathrm{acs}}

, unless stated otherwise. I2I-ReID results are reported on Market-1501 and T2I-ReID results on RSTPReID.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Retrieval Top-1 Error-rate distribution across entropy and cosine distance. Curves computed on the Market-1501 [ 52 ] dataset for I2I-ReID and RSTPReID [ 53 ] for T2I-ReID.

Panel of Figure 6 — Figure 6: Understanding the nature of the embedding space between a query and gallery, and a theoretical interpretation of Figure 5 where error rate is a surrogate for $e(r)$ .

Figure 7: Prompt-bank construction pipeline. (a) A VLM is used to caption unlabeled pedestrian images and extract recurring appearance cues. These cues are grouped into semantic categories and cleaned to remove duplicate or near-identical phrases. (b) The cleaned attribute suffixes are inserted into a shared template to form the final prompt bank $\mathcal{P}=\{s_{\ell}\}_{\ell=1}^{L}$ , together with a neutral prompt $s_{0}$ used for prompt centering.

TTA Method	Market-1501 (Corrupted)
	Gaussian Noise		Poisson Noise		Defocus Blur		Motion Blur		Brightness
	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1
No-adapt (Source) [NeurIPS’24]	24.2	50.7	14.5	37.5	13.7	35.7	8.9	26.1	33.5	60.9
TENT [ wang2020tent ] [ICLR’21]	23.4	48.0	9.0	22.6	11.5	34.2	6.9	19.9	29.7	55.3
SHOT [ 27 ] [ICML’20]	23.8	51.1	15.7	36.3	13.2	33.6	1.9	4.9	28.5	55.5
SAR [ 29 ] [ICLR’23]	26.6	52.2	16.8	38.8	14.2	35.0	8.7	24.3	31.8	58.1

Appendix

Table 5: Image Corruptions and Noise. Comparison of TTA methods on I2IReID using a PLIP [ 55 ] backbone. Results are reported on Market-1501 under five corruptions (mAP / R1).

Prompt-bank source	MSMT17		Market-1501		CUHK-PEDES		RSTP-ReID
Prompt-bank source	mAP	R1	mAP	R1	mAP	R1	mAP	R1
No-adapt source model	10.5	34.2	35.0	62.6	46.4	55.9	23.1	34.5
Generic prompt bank	17.8	36.3	45.9	70.1	48.7	56.1	27.1	36.1
Market-1501 mined bank	18.2	38.5	46.5	71.1	49.6	58.4	27.6	37.6
MSMT17 mined bank	18.4	39.1	46.5	70.1	49.2	57.6	26.9	37.0
CUHK-PEDES mined bank	17.9	36.5	45.4	69.7	49.4	58.1	27.6	37.5

Appendix

Table 6: Effect of attribute prompt-bank source. We evaluate PRAISE using the same fixed hyperparameters while varying only the source used to construct the attribute prompt bank. The manual bank is hand-designed, while the remaining banks are mined from the corresponding ReID datasets using the same label-free captioning and cleaning protocol. Results show that PRAISE is not tied to a specific prompt-bank source and remains stable across transferable pedestrian attribute vocabularies.

Prompt-bank variant	Market-1501		RSTP-ReID
Prompt-bank variant	mAP	R1	mAP	R1
No-adapt source model	35.0	62.6	23.1	34.5
Full prompt bank	46.5	71.1	27.6	37.6
w/o upper-body color	45.8	68.1	26.1	35.2
w/o lower-body color	45.9	69.3	27.0	36.9
w/o upper-body clothing	46.0	69.5	27.2	37.4

Appendix

Table 7: Attribute-category ablation. We evaluate PRAISE using the same Market-1501 mined prompt bank and fixed hyperparameters, while removing one group of attributes at a time. This isolates which attribute categories contribute most to the query-conditioned ACS geometry.

TTA Method	I2I-ReID
	Train: MSMT17				Train: Market-1501
	Test: MSMT17		Test: Market-1501		Test: MSMT17		Test: Market-1501
	mAP	R1	mAP	R1	mAP	R1	mAP	R1
No-adapt (Source)	57.4	79.9	37.8	63.3	21.0	46.9	83.6	92.6
TENT [ wang2020tent ] [ICLR’21]	55.2	75.4	35.2	60.5	18.6	44.2	82.1	91.1
SHOT [ 27 ] [ICML’20]	49.7	68.4	32.1	57.6	14.6	41.2	80.6	89.8

Appendix

Table 8: Comparison of TTA methods on I2I-ReID benchmarks using a source-trained CLIP-ReID [ 26 ] image encoder. All methods are evaluated under the same PRAISE-compatible cosine-retrieval protocol, with fixed gallery embeddings and no target identity labels.

TTA Method	T2I-ReID
	Train: CUHK-PEDES				Train: RSTP-ReID
	Test: CUHK		Test: RSTP		Test: CUHK		Test: RSTP
	mAP	R1	mAP	R1	mAP	R1	mAP	R1
No-adapt (Source) [NeurIPS’24]	64.2	72.7	41.2	54.9	33.5	36.8	47.9	62.1
TENT [ wang2020tent ] [ICLR’21]	64.0	72.8	41.0	53.4	33.1	36.4	47.5	62.0
SHOT [ 27 ] [ICML’20]	64.5	72.9	41.3	54.7	33.8	37.1	48.3	62.5

Appendix

Table 9: Comparison of TTA methods on T2I-ReID benchmarks using source-trained RDE [ qin2024noisy ] backbones. All methods are evaluated under the same PRAISE-compatible cosine-retrieval protocol, with fixed gallery embeddings and no target identity labels.

Figure 8: Gallery-size scalability of PRAISE on MSMT17 with PLIP. Peak GPU-memory and adaptation-latency increases are reported relative to the 1k-gallery setting. The gallery is increased from 1,000 to the full 82,161 images while retaining the default adaptation batch size of 32. Even at the full gallery, peak memory increases by only

9.0\%

, while amortized per-query latency increases by

15.5\%

Figure 9: Sensitivity to initial retrieval quality on Market-1501. Queries are grouped by the initial rank of the first valid cross-camera match under the frozen PLIP baseline. Bars show the percentage of queries whose first-correct rank improves, remains unchanged, or degrades after PRAISE. The line reports mean

\Delta\mathrm{RR}

, where positive values indicate improved retrieval.

$\lambda_{\mathrm{att}}$	Market-1501	RSTPReID
0.0	36.8/65.0	23.5/35.4
0.1	39.2/60.3	24.5/32.9
0.2	38.9/60.1	24.3/32.5
0.5	37.5/58.7	23.7/30.8
1.0	35.1/54.5	23.1/28.3

Appendix

Table 10: Sensitivity to the direct attribute-fusion weight

\lambda_{\mathrm{att}}

. Results are reported as mAP/R1.

Figure 10: NMI comparison across methods.

Figure 11: Some success and failure cases for I2I-ReID from Market-1501 [ 52 ] dataset.

Figure 12: Some success and failure cases for T2I-ReID from RSTPReID [ 53 ] dataset.

Person Re-Identification

Person re-identification matches the same individual across camera views, times, clothing, and sensing modalities despite changes in appearance.

See all 92 papers

ALADIN:Attribute-Language Distillation Network for Person Re-Identification
Mar 23, 2026Wang Zhou, Boran Duan, Haojun Ai +2Person Re-Identification Cross-Modal Knowledge Distillation
Wuhan University No.299 Bayi Road, 430072, PR China
Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification
Mar 20, 2026Kunlun Xu, Haotong Cheng, Jiangmeng Li +2Person Re-Identification Continual Learning for VLMs
Wangxuan Institute of Computer Technology, Peking University, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China
ReText: Text Boosts Generalization in Image-Based Person Re-identification
Feb 5, 2026Timur Mamedov, Karina Kvanchiani, Anton Konushin +1Person Re-Identification
Tevian, Moscow, Russia · Lomonosov Moscow State University

Mar 23, 2026J/K move · Enter open · S save

ALADIN:Attribute-Language Distillation Network for Person Re-Identification

Wang Zhou, Boran Duan, Haojun Ai, Ruiqi Lan+1

Wuhan University No.299 Bayi Road, 430072, PR China

Person Re-Identification Cross-Modal Knowledge Distillation Vision-Language Alignment Visual Prompt Learning

Open paper AskPDF arXiv

Recent vision-language models such as CLIP provide strong cross-modal alignment, but current CLIP-guided ReID pipelines rely on global features and fixed prompts. This limits their ability to capture fine-grained attribute cues and adapt to diverse appearances. We propose ALADIN, an attribute-language distillation network that distills knowledge from a frozen CLIP teacher to a lightweight ReID student. ALADIN introduces fine-grained attribute-local alignment to establish adaptive text-visual correspondence and robust representation learning. A Scene-Aware Prompt Generator produces image-specific soft prompts to facilitate adaptive alignment. Attribute-local distillation enforces consistency between textual attributes and local visual features, significantly enhancing robustness under occlusions. Furthermore, we employ cross-modal contrastive and relation distillation to preserve the inherent structural relationships among attributes. To provide precise supervision, we leverage Multimodal LLMs to generate structured attribute descriptions, which are then converted into localized attention maps via CLIP. At inference, only the student is used. Experiments on Market-1501, DukeMTMC-reID, and MSMT17 show improvements over CNN-, Transformer-, and CLIP-based methods, with better generalization and interpretability.

Abstract

Person Re-Identification

Explore similar work

ALADIN:Attribute-Language Distillation Network for Person Re-Identification

Vision-Language Attribute Disentanglement and Reinforcement for Lifelong Person Re-Identification

ReText: Text Boosts Generalization in Image-Based Person Re-identification