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

Wang Zhou; Boran Duan; Haojun Ai; Ruiqi Lan; Ziyue Zhou

cs.CVMar 23, 2026

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

Authors: Wang Zhou, Boran Duan, Haojun Ai, Ruiqi Lan, Ziyue Zhou

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

Abstract

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.

Figures & tables

Figure 1: Overview of the proposed attribute–language distillation framework. Given an input image, an MLLM generates structured attribute descriptions, which are parsed into semantic groups and encoded by the frozen CLIP text encoder with scene-aware soft prompts. The CLIP image encoder provides attention maps and attribute vectors that guide multi-level distillation. These semantic cues are transferred to a lightweight student ReID backbone, enabling fine-grained attribute alignment and robust identity representation.

Figure 2: Overall pipeline of the proposed multi-level attribute–language distillation framework. A frozen CLIP encoder provides global embedding $\mathbf{g}^{t}$ , attention maps, attribute tokens $\mathbf{T}^{t}_{a}$ and local attribute tokens $\mathbf{T}^{t}_{a}=\{T^{t}_{[x_{1}]},\dots,T^{t}_{[x_{i}]}\}$ . The student network extracts raw features $\mathbf{f}^{s}_{\text{raw}}$ , which are supervised through three branches: (1) local alignment loss $\mathcal{L}_{\text{local}}$ , (2) attribute semantic loss $\mathcal{L}_{\text{attr}}$ , and (3) global feature distillation loss $\mathcal{L}_{\text{feat}}$ guided by $\mathbf{g}^{t}$ . During inference, the Batch Normalization Neck feature $\mathbf{g}^{s}$ is used for cosine-similarity retrieval, yielding refined ranking performance.

Backbone	Method	Ref.	Market1501		MSMT17		DukeMTMC
Backbone	Method	Ref.	mAP	R-1	mAP	R-1	mAP	R-1
ResNet50	IBN-Net [ 22 ]	ECCV’18	88.2	95.0	-	-	79.1	90.1
	OSNet [ 23 ]	ICCV’19	84.9	94.8	52.9	78.7	73.5	88.6
	CDNet [ 24 ]	CVPR’21	86.0	95.1	54.7	78.9	76.8	88.6
	CAL [ 25 ]	ICCV’21	87.0	94.5	56.2	79.5	76.4	87.6
	ALDER [ 26 ]	TIP’21	88.9	95.6	59.1	82.5	78.9	89.9

Table 1: Comparison with state-of-the-art person ReID methods on three benchmarks. Best results are in bold .All the results are without Re-ranking.

Method	mAP (%)	Rank-1 (%)
Baseline (ID+Triplet)	88.2	95.0
+ Global Distill	88.6	95.3
+ Mid-layer Distill	88.7	95.4
+ Attribute Semantic	89.0	95.6
+ Local Align	89.2	95.7
+ Attribute CE	89.4	95.8

Table 2: Ablation on Market-1501 (single-query, without re-ranking). We add components progressively.

Setting	Noise Type	Ratio	Market-1501
Setting	Noise Type	Ratio	mAP	R1
ALADIN(Ours)	None	0.0	89.4	96.0
Filter- $r$	Drop (field removal)	0.2	89.2	95.8
Filter- $r$	Drop (field removal)	0.4	88.8	95.2
Filter- $r$	Drop (field removal)	0.6	88.4	95.0
Filter- $r$	Drop (field removal)	0.8	88.2	94.9

Table 3: Impact of MLLM attribute noise on ReID performance. “Drop” removes (filters) attribute fields by replacing them with unknown ; “Wrong” flips attribute values within the same key.

Error Type	$p$	mAP	R1
Color error only	0.2	89.0	95.5
Category error only (top/bottom/shoes)	0.2	88.8	95.3
Existence error only (carrying/accessory)	0.2	88.7	95.1
Mixed (all above)	0.2	88.0	94.4

Table 4: Breakdown of wrong-attribute types at a fixed wrong ratio

p=0.2

on Market-1501.

Variant	Prompt Type	mAP	R1
w/o Soft Prompt	None	88.7	95.5
Random Prompt (control)	Random, not trained	88.5	95.4
Fixed Soft Prompt	Learnable, global	89.0	95.7
Scene-Aware Soft Prompt (Ours)	Conditioned, per-sample	89.4	96.0

Table 5: Ablation of soft prompt strategies on Market-1501. Fixed prompts are shared across samples; scene-aware prompts are conditioned on the student feature.

Prompt Length $N$	mAP	R1
2	89.1	95.8
4	89.4	96.0
8	89.5	96.2

Table 6: Effect of soft prompt length

N

(number of soft tokens) on Market-1501.

Datasets	Models	Rank-1 (%)	Rank-5 (%)	mAP (%)
M $\rightarrow$ D	ResNet-50	45.02	61.40	26.43
M $\rightarrow$ D	AMD [ 14 ]	47.08	62.61	28.41
	Ours	48.20	63.95	29.42
D $\rightarrow$ M	ResNet-50	53.15	71.26	24.48
D $\rightarrow$ M	AMD [ 14 ]	54.48	71.59	25.48
	Ours	55.90	72.30	26.55

Table 7: Cross-domain evaluation results of ResNet-50 and AMD. M→D: trained on Market-1501 and tested on DukeMTMC-ReID; D→M: reverse setting.

Figure 3: Qualitative comparison of the same retrieval batch under baseline and our method. Both rows display identical gallery images but with distinct attention patterns. The baseline exhibits uniform, global attention across the entire person region, lacking fine-grained discrimination. In contrast, our method is designed to focus on semantically meaningful local attributes,such as clothing texture, color, and accessories, enabling more discriminative matching and improved interpretability. This shift from global to attribute-aware attention contributes to higher retrieval accuracy in challenging cases.

Person Re-Identification

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

See all 84 papers

Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification
Mar 14, 2026Jiachen Li, Xiaojin Gong, Dongping ZhangPerson Re-Identification Vision-Language Alignment
College of Information Science and Electronic Engineering, Zhejiang University · College of Information Engineering, China Jiliang University
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
DiffReID: Discriminative Diffusion Model for Object Re-Identification
Sep 29, 2026Yingquan Wang, Pingping Zhang, Dong Wang +1Representation Learning Query Re-Identification
School of Information and Communication Engineering, Dalian University of Technology · School of Future Technology, Dalian University of Technology

Mar 14, 2026J/K move · Enter open · S save

Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification

Jiachen Li, Xiaojin Gong, Dongping Zhang

College of Information Science and Electronic Engineering, Zhejiang University · College of Information Engineering, China Jiliang University

Person Re-Identification Vision-Language Alignment

Open paper AskPDF arXiv

Domain Generalized person Re-identification (DG Re-ID) is a challenging task, where models are trained on source domains but tested on unseen target domains. Although previous pure vision-based models have achieved significant progress, the performance remains further improved. Recently, Vision-Language Models (VLMs) present outstanding generalization capabilities in various visual applications. However, directly adapting a VLM to Re-ID shows limited generalization improvement. This is because the VLM only produces with global features that are insensitive to ID nuances. To tacle this problem, we propose a CLIP-based multi-grained vision-language alignment framework in this work. Specifically, several multi-grained prompts are introduced in language modality to describe different body parts and align with their counterparts in vision modality. To obtain fine-grained visual information, an adaptively masked multi-head self-attention module is employed to precisely extract specific part features. To train the proposed module, an MLLM-based visual grounding expert is employed to automatically generate pseudo labels of body parts for supervision. Extensive experiments conducted on both single- and multi-source generalization protocols demonstrate the superior performance of our approach. The implementation code will be released at https://github.com/RikoLi/MUVA.

Abstract

Person Re-Identification

Explore similar work

Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification

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

DiffReID: Discriminative Diffusion Model for Object Re-Identification