DADKC: Distribution-Aware Adaptive Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification

Yan Zhang; Minghao Long; Meiqi Wan; Yan Yang; Xin Yuan; Wei Wang; Xiao Wang

cs.CVOct 31, 2026

DADKC: Distribution-Aware Adaptive Differentiated Knowledge Consolidation for Cloth-Hybrid Lifelong Person Re-identification

Authors: Yan Zhang, Minghao Long, Meiqi Wan, Yan Yang, Xin Yuan, Wei Wang, Xiao Wang

Organizations: School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan 430065, China · State Key Laboratory of Robotics and Intelligent Systems, Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China · China University of Chinese Academy of Sciences, Beijing 100049, China · Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System, Wuhan University of Science and Technology, Wuhan 430065, China

Abstract

Lifelong Person Re-identification (LReID) aims to match the same person using sequentially collected data. In practical long-term deployments, however, the model must handle both domain shifts across cameras and scenes, and irregular clothing changes of the same identity. This gives rise to the Cloth-Hybrid Lifelong Person Re-identification (CH-LReID) task, which demands balancing cloth-relevant and cloth-irrelevant knowledge in streaming data. While existing LReID methods have explored replay-based or distillation-based anti-forgetting strategies, they typically fail to sufficiently model stage-aware domain distributions and cross-stage feature continuity under exemplar-free training. To address this problem, we propose a Distribution-Aware Adaptive Differentiated Knowledge Consolidation (DADKC) framework. The core idea is to enhance the existing Differentiated Knowledge Consolidation (DKC) paradigm with adaptive distribution modeling. To this end, three lightweight modules are designed: (1) an Anchor Pool Adapter (APA) to retrieve learnable anchor cues for refining identity features; (2) a Phase Path Mixer (PPM) to perform stage-aware expert routing for integrating domain-specific distribution paths; and (3) a Stage Bridge Alignment (SBA) module to regularize the continuity of base, anchor, and path representations between adjacent stages. Extensive experiments on the CH-LReID and LReID benchmarks show that DADKC gains 4.1%/4.5% mAP/R1 on Order-1 and 3.6%/3.4% mAP/R1 on Order-2 over the DKC baseline, and lowers weighted forgetting from 2.4%/1.7% to 1.8%/1.4%, while storing neither historical images nor per-image features. The code is available at https://github.com/fire667/DADKC.

Figures & tables

Figure 1: Paradigm comparison of LReID methods: rehearsal-based methods replay exemplars, rehearsal-free methods avoid raw samples but may lose cross-stage continuity, and DADKC performs distribution-aware differentiated consolidation.

Figure 2: Overall framework of DADKC. A DKC-style backbone extracts base and uncertainty representations; APA retrieves anchor cues, PPM routes through stage-aware experts, and SBA aligns current and previous-stage base, anchor, and path features. The final model uses accumulated experts and an unknown-domain embedding at test time.

Method	Cloth-Consistent												Cloth-Changing						Average
	Market		LTCC		PRCC		MSMT17		CUHK03		Avg.		LTCC		PRCC		Avg.		mAP	R1
	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1
Joint	64.1	82.5	42.6	62.1	94.6	98.7	18.4	40.8	44.4	46.4	52.8	66.1	10.1	23.0	32.7	33.8	21.4	28.4	43.8	55.3
SFT	28.5	52.0	28.5	49.3	92.5	97.3	7.0	19.6	44.0	45.6	40.1	52.8	6.9	15.3	21.8	21.2	14.4	18.3	32.7	42.9
AKA (CVPR 2021) [ 10 ]	48.0	69.5	25.4	45.1	88.1	93.3	4.2	12.0	31.2	31.2	39.4	50.2	6.5	12.8	26.5	26.7	16.5	19.8	32.8	41.5
PatchKD (ACM MM 2022) [ 15 ]	68.0	85.5	30.8	54.7	93.5	96.5	5.7	15.6	33.2	32.9	46.2	57.0	7.2	17.9	26.1	26.0	16.7	22.0	37.8	47.0

Table 1: Order-1 performance. DKC-reported values are from DKC [ 1 ] ; DKC-controlled and DADKC are common-code five-seed means.

Method	Cloth-Consistent												Cloth-Changing						Average
	MSMT17		PRCC		Market		CUHK03		LTCC		Avg.		PRCC		LTCC		Avg.		mAP	R1
	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1	mAP	R1
Joint	18.4	40.8	94.6	98.7	64.1	82.5	44.4	46.4	42.6	62.1	52.8	66.1	32.7	33.8	10.1	23.0	21.4	28.4	43.8	55.3
SFT	2.4	9.1	84.5	95.9	16.9	37.2	8.2	8.8	37.9	57.6	30.0	41.7	21.9	24.3	9.1	20.9	15.5	22.6	25.8	36.3
AKA (CVPR 2021) [ 10 ]	15.0	32.4	93.0	97.0	26.6	51.0	13.4	11.6	43.4	64.0	38.3	51.2	28.2	27.7	8.4	15.8	18.3	21.8	32.6	42.8
PatchKD (ACM MM 2022) [ 15 ]	22.6	47.9	95.0	98.5	38.2	64.2	21.8	23.1	45.8	70.0	44.7	60.7	28.1	27.6	9.3	21.7	18.7	24.7	37.3	50.4

Table 2: Order-2 performance. DKC-reported values are from DKC [ 1 ] ; DKC-controlled and DADKC are common-code five-seed means.

Method	Order-3		Order-4		Average
Method	mAP	R1	mAP	R1	mAP	R1
CRL (WACV 2021) [ 33 ]	40.5	50.8	37.6	49.2	39.1	50.0
AKA (CVPR 2021) [ 10 ]	40.8	50.8	38.6	49.4	39.7	50.1
PatchKD (ACM MM 2022) [ 15 ]	43.7	53.7	43.2	54.1	43.5	53.9
LSTKC (AAAI 2024) [ 24 ]	50.0	63.1	49.6	62.1	49.8	62.6
DKP (CVPR 2024) [ 23 ]	51.8	64.1	51.3	63.2	51.6	63.7

Table 3: Traditional cloth-consistent LReID results. Baselines are from DKC [ 1 ] ; DKC-controlled and DADKC are five-seed means.

Method	Same (5 evals.)	Cross (2 evals.)	Weighted all
	AF(mAP/R1) $\downarrow$	AF(mAP/R1) $\downarrow$	AF(mAP/R1) $\downarrow$
AKA (CVPR 2021) [ 10 ]	12.9/13.8	5.0/6.8	10.6/11.8
LSTKC (AAAI 2024) [ 24 ]	15.9/13.0	7.5/9.3	13.5/11.9
USP (TCSVT 2024) [ 27 ]	13.6/11.8	5.6/7.3	11.3/10.5
PatchKD (ACM MM 2022) [ 15 ]	6.3/6.3	4.5/5.4	5.8/6.0
DKP (CVPR 2024) [ 23 ]	7.8/6.2	3.0/4.8	6.4/5.8

Table 4: Order-1 AF results. Same and Cross average five and two evaluations; Weighted all uses

5:2

weighting. Lower is better.

Variant	Purpose	Avg. mAP	Avg. R1
DKC-reported	published reference	44.2	55.8
DKC + matched MLP	capacity control	44.5	56.0
DKC + generic feature KD	alignment control	44.6	56.3
DKC + generic MoE	routing control	44.8	56.4
DKC + APA	anchor only	44.7	56.3
DKC + PPM	path only	44.9	56.5

Table 5: Order-1 component and capacity analysis. DKC-reported is the published reference; other rows are common-pipeline five-seed variants.

Parameter	Values	Avg. mAP	Avg. R1
$M$	$64/128^{\dagger}/256$	45.3 / 45.7 / 45.4	57.1 / 57.6 / 57.2
$K$	$4/8^{\dagger}/12$	45.1 / 45.7 / 45.5	56.9 / 57.6 / 57.3
$E$	$3/5^{\dagger}/7$	45.4 / 45.7 / 45.3	57.2 / 57.6 / 57.1
$\lambda_{apa}=\lambda_{ppm}$	$.05/.10^{\dagger}/.20$	45.3 / 45.7 / 45.4	57.2 / 57.6 / 57.3
$\lambda_{sba}$	$.005/.01^{\dagger}/.02$	45.4 / 45.7 / 45.5	57.3 / 57.6 / 57.4

Table 6: Order-1 sensitivity. † marks the default used by every main table.

Method	Params	GFLOPs	Train GB	ms/image	Growth/stage
DKC	28.7M	4.11	7.2	4.8	0
Matched MLP	31.1M	4.45	7.8	5.2	0
DADKC	31.2M	4.48	9.9	5.5	0.48M

Table 7: Resource profile after five stages. Training memory includes the frozen SBA teacher; latency is single-image inference after warm-up.

Figure 3: t-SNE visualization of learned representations on training Order-1 . Different colors represent different identities, and

\triangle

\bigcirc

denote two randomly sampled identities with different clothes.

Figure 4: Performance tendency at different training stages on Order-1 .

Exemplar-Free Class-Incremental Learning

See all 9 papers

SCORE: SubDistribution-aware Collaborative Knowledge Reinforcing for Cloth-Hybrid Lifelong Person Re-Identification
Sep 11, 2026Kunlun Xu, Liangyu Ma, Jiangmeng Li +4Prototype-Based Learning Catastrophic Forgetting
Wangxuan Institute of Computer Technology, Peking University, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · Intelligent Science & Technology Academy of CASIC, Beijing 100041, China
AdalReID: Auto-selected Knowledge Adapters for Lifelong Person Re-identification
May 29, 2024Xuelin Qian, Ruiqi Wu, Zhongling Huang +2Person Re-Identification Adapter-Based Continual Learning
Northwestern Polytechnical University
Beyond Discrete Samples: High Information Density Replay for Efficient Lifelong Person Re-Identification
Aug 3, 2025Mingyu Wang, Haojie Liu, Zhiyong Li +3Replay-Based Continual Learning Dataset Condensation
College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, China

Sep 11, 2026J/K move · Enter open · S save

SCORE: SubDistribution-aware Collaborative Knowledge Reinforcing for Cloth-Hybrid Lifelong Person Re-Identification

Kunlun Xu, Liangyu Ma, Jiangmeng Li, Xin Tong+3

Wangxuan Institute of Computer Technology, Peking University, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · Intelligent Science & Technology Academy of CASIC, Beijing 100041, China

Prototype-Based Learning Catastrophic Forgetting Person Re-Identification

Open paper AskPDF arXiv

Lifelong Person Re-Identification (LReID) aims to train a unified person retrieval model from a non-stationary data stream. Existing LReID methods mainly focus on scenarios where the clothing of each person is consistent. Recently, the Cloth-Hybrid LReID (CH-LReID) where cloth-consistent and cloth-changing data alternately occur, has emerged as a more practical and challenging scenario. Due to the conflict between clothing-relevant and clothing-irrelevant knowledge, the well-known catastrophic forgetting problem is significantly exacerbated in this task. To address this issue, we propose a SubDistribution-aware COllaborative Knowledge REinforcing (SCORE) framework, where our key idea is explicitly modeling the intra-identity diversity to continually consolidate distinct cloth-consistent and cloth-changing knowledge. Specifically, an Adaptive SubDistribution Modeling mechanism is developed, where a set of distributional subprototypes is assigned to each identity to capture the intra-identity diversity, improving the compatibility between cloth-consistent and cloth-changing knowledge. Then, a Distributional Knowledge Reinforcement scheme is introduced, where the knowledge of old distributional subprototypes is retained in the new ones by a collaborative aligning mechanism. Extensive experiments show that our SCORE achieves the state-of-the-art performance. Our code is available at https://github.com/zhoujiahuan1991/ECCV2026-SCORE

Abstract

Exemplar-Free Class-Incremental Learning

Explore similar work

SCORE: SubDistribution-aware Collaborative Knowledge Reinforcing for Cloth-Hybrid Lifelong Person Re-Identification

AdalReID: Auto-selected Knowledge Adapters for Lifelong Person Re-identification

Beyond Discrete Samples: High Information Density Replay for Efficient Lifelong Person Re-Identification