PRDP: Person Reidentification With Dirty and Poor Data.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Person reidentification (ReID) models struggle with data issues like mislabeled images (dirty quality) and insufficient samples for some identities (FewIDs).
• Training ReID models with such data leads to poor generalization and reduced performance.

Purpose of the Study:

• To propose a novel method for simultaneously addressing data quality and quantity problems in person reidentification.
• To enhance the robustness and generalization capabilities of ReID models.

Main Methods:

• A weighted label correction based on cross-entropy (wLCCE) strategy is proposed to handle mislabeled data, classifying errors into point and set label errors.
• A weighted triplet loss (WTL) is introduced to correct these identified label errors.
• A feature simulation based on autoencoder (FSAE) method generates virtual samples for identities with few images (FewIDs).
• An autoencoder (AE)-based simulator transfers identity difference patterns from multi-image identities (MultIDs) to FewIDs for richer feature expressions.

Main Results:

• The proposed methods effectively address both dirty quality and poor quantity data issues in person reidentification.
• Experimental results on Market-1501 and DukeMTMC-reID datasets demonstrate the effectiveness of the approach.
• The methods enable the learning of more robust ReID models.

Conclusions:

• The novel approach successfully tackles critical data challenges in person reidentification.
• The combination of label correction and feature simulation significantly improves ReID model performance.
• This work contributes to developing more reliable person reidentification systems.