Weakly Supervised Tracklet Association Learning With Video Labels for Person Re-Identification

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Supervised person re-identification (re-id) demands extensive manual labeling.
• Unsupervised re-id methods offer lower performance compared to supervised alternatives.
• Existing weakly supervised methods often require auxiliary datasets or struggle with noisy data.

Purpose of the Study:

• To develop a novel weakly supervised person re-identification model using only video labels.
• To address the limitations of existing weakly supervised approaches, particularly their reliance on auxiliary data and handling of noisy tracklets.

Main Methods:

• Formulated a weakly supervised tracklet association learning (WS-TAL) model.
• Introduced an intra-bag tracklet discrimination learning (ITDL) term for pseudo-labeling and feature learning.
• Developed a cross-bag tracklet association learning (CTAL) term to mine tracklet associations between bags.
• Jointly optimized ITDL and CTAL terms for model training.

Main Results:

• Achieved 88.1% rank-1 accuracy on the MARS dataset.
• Achieved 90.3% rank-1 accuracy on the DukeMTMC-VideoReID dataset.
• Significantly surpassed state-of-the-art weakly supervised models and even some fully supervised re-id models.

Conclusions:

• The WS-TAL model effectively learns discriminative features and associations using only video labels.
• This approach offers a practical and high-performing solution for person re-identification with reduced labeling effort.
• The proposed method demonstrates superior performance compared to existing weakly supervised and some supervised re-id techniques.