Collaborative Contrastive Refining for Weakly Supervised Person Search.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Weakly supervised person search trains models using only bounding box annotations, lacking identity labels.
• Clustering algorithms often generate inaccurate pseudo-labels and suffer from imbalanced identity distributions, leading to significant noise.
• Existing methods struggle with label noise and imbalanced data, hindering performance in person search tasks.

Purpose of the Study:

• To propose a novel Collaborative Contrastive Refining (CCR) framework for weakly supervised person search.
• To jointly refine pseudo-labels and the sample-learning process using advanced contrastive strategies.
• To address the challenges of inaccurate pseudo-labels and imbalanced identity distributions in person search.

Main Methods:

• Developed a Collaborative Contrastive Refining (CCR) framework utilizing hybrid contrastive strategies.
• Employed a hybrid contrastive strategy combining visual and contextual cues for pseudo-label refinement.
• Implemented a sample-mining and noise-contrastive strategy to mitigate the impact of imbalanced data by distinguishing positive and noise samples.

Main Results:

• The CCR framework demonstrated superior performance in refining pseudo-labels through hybrid similarity exploration.
• The method effectively distinguished query and noise samples, enhancing the sample-learning process.
• Achieved over 3% mAP improvement on the CUHK-SYSU dataset compared to state-of-the-art weakly supervised methods.

Conclusions:

• The proposed CCR framework significantly enhances weakly supervised person search by improving pseudo-label quality and sample learning.
• CCR achieves state-of-the-art results, surpassing existing weakly supervised methods and rivaling supervised approaches.
• The framework's ability to leverage diverse unlabeled data offers a promising direction for future research in person search.