Deep Graph Metric Learning for Weakly Supervised Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Conventional person re-identification (re-id) relies on costly frame-by-frame instance-level labeling from surveillance videos.
• This manual annotation process across multiple camera views is time-consuming and labor-intensive.
• Existing methods struggle with visual ambiguities and appearance variations inherent in re-id tasks.

Purpose of the Study:

• To develop a weakly supervised person re-id model that eliminates the need for instance-level labeling in training.
• To identify target persons within untrimmed video clips using only video-level identity information.
• To address the challenges posed by missing instance-level labels and ground-truth bounding boxes in weakly supervised settings.

Main Methods:

• Deep Graph Metric Learning (DGML) is proposed to measure consistency between intra-video spatial graphs and distinguish inter-video spatial graphs.
• Weakly Supervised Regularization (WSR) is introduced to embed weak supervision, utilizing video-level labels.
• DGML employs a weak identity loss and a cross-video alignment loss for learning discriminative features.

Main Results:

• The proposed DGML method demonstrates the feasibility of weakly supervised person re-id.
• Extensive experiments validate the effectiveness of DGML, including its application to special cases like bag-to-bag re-id.
• The approach successfully learns discriminative features despite the absence of frame-level annotations.

Conclusions:

• Weakly supervised person re-id is achievable and can significantly reduce annotation costs.
• DGML provides an effective framework for person re-identification using limited supervision.
• The developed method offers a practical solution for large-scale surveillance systems.