Organizations: School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan, Hubei 430065, China · Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System, Wuhan University of Science and Technology, Wuhan 430065, China · School of Computer Science, Wuhan University, Wuhan, Hubei 430072, China · Department of Electrical Engineering and the Institute of Communications Engineering, National Tsing Hua University, and with Electronic and Optoelectronic System Research Laboratories, Industrial Technology Research Institute, Hsinchu, Taiwan 100084, China
With the increasing demand for robust person Re-ID in unconstrained environments, learning from datasets with noisy labels and sparse per-identity samples remains a critical challenge. Existing noise-robust person Re-ID methods primarily rely on loss-correction or sample-selection strategies using softmax outputs. However, these methods suffer from two key limitations: 1) Softmax exhibits translation invariance, leading to over-confident and unreliable predictions on corrupted labels. 2) Conventional sample selection based on small-loss criteria often discards valuable hard positives that are crucial for learning discriminative features. To overcome these issues, we propose the CAlibration-to-REfinement (CARE) method, a two-stage framework that seeks certainty through probabilistic evidence propagation from calibration to refinement. In the calibration stage, we propose the probabilistic evidence calibration (PEC) that dismantles softmax translation invariance by injecting adaptive learnable parameters into the similarity function, and employs an evidential calibration loss to mitigate overconfidence on mislabeled samples. In the refinement stage, we design the evidence propagation refinement (EPR) that can more accurately distinguish between clean and noisy samples. Specifically, the EPR contains two steps: Firstly, the composite angular margin (CAM) metric is proposed to precisely distinguish clean but hard-to-learn positive samples from mislabeled ones in a hyperspherical space; Secondly, the certainty-oriented sphere weighting (COSW) is developed to dynamically allocate the importance of samples according to CAM, ensuring clean instances drive model updates. Extensive experimental results on Market1501, DukeMTMC-ReID, and CUHK03 datasets under both random and patterned noises show that CARE achieves competitive performance.
Figures & tables
Fig. 1 : Illustration of core idea. Left: two challenging instances in feature space, the sample 4 (clean but close to another identity) and the sample 5 (hard sample with occlusion). Right: (a) Original samples with noisy labels; (b) Sample selection methods filter out noisy but informative samples [ 8 , 9 , 10 ] ; (c) DistributionNet uses uncertainty to model features, yet it still confuses similar features between clean and noisy labels [ 4 ] ; (d) Label refinement methods based on softmax may produce the same probabilities for different samples, resulting in incorrectly refurbished labels [ 5 , 6 , 11 , 12 ] ; CARE contains (e) evidential calibration and (f) evidential refinement: (e) calibrates the high evidential instances in the Calibration stage, and (f) will refine the low evidential instances in the Refinement stage. Best viewed in color .
Notation
Shape/Domain
Description
Notation
Shape/Domain
Description
D
{(xi,yi)}i=1N
training dataset
N
scalar
number of samples
C
scalar
number of identity classes
xi
X
i -th input image
yi
{1,…,C}
identity label
zij
scalar
logit score
zi
RC
logit vector
sj
scalar ≥0
smoothing term
Θ
network params
backbone parameters
φ(⋅)
mapping
evidence mapping
υi
R≥0C
evidence vector
κ
scalar >0
concentration offset
TABLE I: Summary of important notations used in the paper.
Fig. 2 : Illustration of the core problems and our solution under 20% random noise. (a) Top row : conventional softmax-based scoring yields over-confident predictions on corrupted labels, while small-loss selection tends to discard informative but hard positives. (b) Bottom row : Our two-stage CARE framework addresses these problems by first calibrating uncertainty to isolate noise in the Calibration stage, then refining with angular metrics to preserve hard positives through soft weighting in the Refinement stage. Simple, noisy, and hard positive samples are marked as , , and , respectively.
Fig. 3 : Illustration of overall CARE framework. Calibration stage: PEC integrates Dirichlet‑informed prediction calibration to break softmax’s translation invariance to mitigate over‑confidence. Refinement stage: EPR, powered by the CAM metric, surpasses small-loss methods in distinguishing clean but hard‑to‑learn samples from mislabeled ones; then COSW dynamically reallocates sample importance to prioritize clean instances over noisy instances.
Datasets
Market1501
DukeMTMC-ReID
CUHK03
Noise
Methods
Venue
Fields
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
10%
SL* [ 33 ]
ICCV19
IC
79.2
91.8
94.8
57.0
68.6
81.7
86.7
50.2
33.9
54.7
65.2
33.5
MMT* [ 45 ]
ICLR20
IC
81.1
93.5
95.8
63.2
73.0
86.1
90.3
59.6
27.2
49.2
60.7
29.7
JoCoR* [ 7 ]
CVPR20
IC
81.5
93.4
96.0
60.8
68.2
81.2
86.3
48.3
28.8
49.2
59.6
29.4
Focal* [ 32 ]
TPAMI20
IC
78.7
91.5
94.8
57.3
68.3
82.1
86.5
50.4
-
-
-
-
UbiW* [ 34 ]
TPAMI21
IC
79.1
91.9
95.0
56.8
68.9
81.8
86.7
49.6
30.1
50.4
60.9
29.5
TABLE II : Performance results with state-of-the-art methods under different random noise ratios (10%, 20%, 30%, and 50%). “IC” denotes the methods designed for the image classification task, whereas “PR” refers to those methods developed for the person Re-ID task. “*” indicates the results are reported from their original papers, while the remaining results are reproduced under the same experimental settings. Best results are in bold .
Datasets
Market1501
DukeMTMC-ReID
CUHK03
Noise
Methods
Venue
Fields
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
10%
SL* [ 33 ]
ICCV19
IC
80.2
92.0
94.8
59.1
70.4
82.9
87.3
50.6
32.6
55.9
64.9
31.5
JoCoR* [ 7 ]
CVPR20
IC
80.9
92.2
95.0
60.2
71.4
83.6
87.8
52.5
31.4
55.2
64.2
30.8
ACT [ 35 ]
ACMMM24
IC
88.2
96.2
97.6
64.5
72.7
85.6
89.4
53.4
32.9
52.3
61.4
30.4
SED [ 46 ]
ECCV24
IC
88.0
96.1
97.5
63.8
72.8
85.6
89.3
53.2
32.7
52.1
61.1
30.1
CA2C [ 47 ]
ICCV25
IC
88.3
96.2
97.6
64.9
73.1
85.8
89.6
55.3
33.0
52.5
61.9
31.0
TABLE III : Performance results with state-of-the-art methods under different patterned noise ratios (10% and 20%). “IC” denotes the methods designed for the image classification task, whereas “PR” refers to those methods developed for the person Re-ID task. “*” indicates the results are reported from their original papers, while the remaining results are reproduced under the same experimental settings. Best results are in bold .
Datasets
Market1501
DukeMTMC-ReID
CUHK03
Noise Type
Noise
Stages
Net
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Random
0%
CORE (S1)
1
90.0
97.1
98.4
69.8
76.3
87.7
91.0
58.8
40.6
60.3
70.1
39.8
CARE (S1)
1
90.8
97.6
98.4
71.0
78.1
88.4
91.2
60.3
44.6
63.4
72.1
42.3
CORE (S1+S2)
2
90.8
97.6
98.6
72.0
78.4
88.1
91.3
62.1
43.6
64.4
73.1
43.4
CARE (S1+S2)
2
90.7
97.7
98.5
72.5
78.4
88.2
91.3
61.2
45.2
64.5
73.9
43.1
20%
CORE (S1)
1
86.4
95.4
97.0
57.7
67.6
81.3
85.1
47.0
24.9
43.3
54.1
23.7
TABLE IV : Performance results in different stages under different random and patterned noise ratios on three person Re-ID datasets. “S1” and “S2” correspond to the Calibration stage and the Refinement stage, respectively. “Net” denotes the number of networks used for training. Best results are in bold .
Datasets
Market1501
DukeMTMC-ReID
CUHK03
Noise Type
Stages
LENLL
LKL(Dir)
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Random
Calibration
✓
85.4
95.2
97.1
56.8
67.5
81.6
85.6
46.8
19.4
37.0
46.8
19.2
✓
✓
87.1
96.0
97.6
60.9
69.7
82.8
87.7
50.5
25.3
42.5
53.2
24.0
LWCE (w/o Noisy )
LWCE (w/ Noisy )
LWKL
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Calibration + Refinement
✓
88.0
96.2
97.5
64.0
73.0
85.6
89.3
55.0
28.3
47.1
57.1
27.2
✓
89.0
97.0
98.0
66.0
73.7
86.0
90.1
55.5
29.2
47.3
58.1
27.7
TABLE V : Performance results with different losses under 20% random and patterned noises on three person Re-ID datasets. Best results are in bold .
Datasets
Market1501
DukeMTMC-ReID
CUHK03
Noise Type
Noise
Batch Size
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Rank-1
Rank-5
Rank-10
mAP
Random
0%
8
79.0
92.1
94.8
45.2
46.5
64.9
71.2
30.0
22.9
44.0
54.4
24.6
16
90.6
97.2
97.9
69.7
73.9
86.0
89.8
57.1
45.9
65.9
74.5
45.0
32
90.7
97.7
98.5
72.5
78.4
88.2
91.3
61.2
45.2
64.5
73.9
43.1
64
90.1
97.5
98.2
71.3
77.4
87.6
90.9
59.5
35.5
56.9
66.4
36.2
128
89.8
97.2
97.9
68.6
75.8
87.3
90.4
58.1
30.2
48.6
58.6
30.0
TABLE VI : Performance results with different batch sizes (8, 16, 32, 64, 128) under random and patterned noise ratios on three person Re-ID datasets. “Batch Size” denotes the number of samples in each training iteration. Best results are in bold .
Fig. 4 : Ablation results of hyperparameters ( λ,α,β ) and compute overhead under 20% random noise ratio. Each row shows results on one person Re-ID dataset: Market1501 (top), DukeMTMC-ReID (middle), and CUHK03 (bottom). From left to right, the columns detail: the ablation studies for hyperparameters λ , α , and β , followed by the compute overhead.
Fig. 5 : Ablation results of hyperparameters ( λ,α,β ) and compute overhead under 20% patterned noise ratio. Each row shows results on one person Re-ID dataset: Market1501 (top), DukeMTMC-ReID (middle), and CUHK03 (bottom). From left to right, the columns detail: the ablation studies for hyperparameters λ , α , and β , followed by the compute overhead.
Fig. 6 : The t-SNE visualizations of 10 identities for CORE and CARE under 20% random and patterned ratios on three person Re-ID benchmarks. The rows correspond to datasets (top: Market1501; middle: DukeMTMC-ReID; bottom: CUHK03). The columns correspond to CORE (random), CARE (random), CORE (patterned), and CARE (patterned). Circles denote correctly labeled samples and crosses denote mislabeled ones. Vc and Va quantify cluster compactness and mislabel dispersion, respectively.
Fig. 7 : Top-10 retrieved results on three person Re-ID datasets, i.e. , Rows 1–2: Market1501; rows 3–4: DukeMTMC-ReID; rows 5–6: CUHK03. Each row shows one query ( left ) and its top-10 retrieved results. The green and red bounding boxes denote correct and wrong retrieval results, respectively.
School of Software, Nanjing University of Information Science and Technology, Nanjing 210044, China · School of Computer Science, Wuhan University, Wuhan 430072, China · School of Computer Science, Nanjing University of Information Science and Technology, Nanjing 210044, China +5
University of Science and Technology of China, China · Anhui Province Key Laboratory of Digital Security, China · The Chinese University of Hong Kong, China