Unsupervised cross-domain person re-identification (re-ID) is difficult because a model trained on one camera domain must generalise to another. The domain shift produces pseudo-labels that are noisy, and when those labels are used in iterative adaptation, the errors can be reinforced rather than corrected.
The new arXiv paper, 'Confidence-Guided Pseudo-Label Refinement for Unsupervised Cross-Domain Person Re-Identification,' proposes to refine pseudo-labels using confidence. The idea is to reduce the influence of low-confidence labels during adaptation, addressing the noise-amplification problem. The abstract contrasts this with existing clustering-based methods, which 'often rely on limited…' — the sentence is cut off in the source.
The provided abstract is truncated, so the exact refinement mechanism and any experimental results are not available in the source. This digest therefore reports only the problem statement and the proposed direction.