From Calibration to Refinement: Seeking Certainty via Probabilistic Evidence Propagation for Noisy-Label Person Re-Identification

约 417 字大约 1 分钟

2026-03-27

News

The paper "CARE" has been accepted for publication in IEEE TMM 2026.

Comments: Accepted by IEEE TMM 2026
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2602.23133 [cs.CV]

Introduction

Limitations of Existing Methods:

Left:

Two challenging instances in feature space: Sample 4 (clean but close to another identity) and Sample 5 (hard sample with occlusion).

Right:

(a) Original samples with noisy labels;
(b) Sample selection methods filter out noisy but informative samples;
(c) DistributionNet uses uncertainty to model features, yet it still confuses similar features between clean and noisy labels;
(d) Label refinement methods based on softmax may produce the same probabilities for different samples, resulting in incorrectly refurbished labels;

CARE contains (e) evidential calibration and (f) evidential refinement:

Conventional Methods (Top Row):

Over-confidence: Traditional softmax-based scoring yields over-confident predictions even on corrupted labels.
Data Loss: Small-loss selection methods tend to discard informative but hard positive samples, leading to suboptimal training.

CARE Framework (Bottom Row):

The two-stage CARE framework addresses these issues as follows:

Stage 1: Calibration: Calibrates uncertainty to effectively isolate noise.
Stage 2: Refinement: Utilizes angular metrics and soft weighting to preserve hard positives.

Sample Legend:

Detailed Two-Stage Paradigm:

Calibration Stage: PEC (Prototypical Evidential Calibration) integrates Dirichlet-informed prediction calibration to break softmax's translation invariance and mitigate over-confidence.
Refinement Stage: EPR (Evidential Prediction Refinement), powered by the CAM metric, surpasses small-loss methods in distinguishing clean but hard-to-learn samples from mislabeled ones; then COSW (Contribution-Oriented Sample Weighting) dynamically reallocates sample importance to prioritize clean instances over noisy instances.