Learning Person Re-Identification Models From Videos With Weak Supervision.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised person re-identification (re-ID) demands extensive frame-level annotations, hindering scalability.
• Unsupervised methods lack the performance of supervised approaches despite eliminating the need for labels.

Purpose of the Study:

• To develop a person re-identification model that utilizes weak video-level supervision, reducing annotation burden.
• To address the performance gap between supervised and unsupervised re-ID methods.

Main Methods:

• Proposed a multiple instance attention learning framework for video person re-identification.
• Framed the task as multiple instance learning, treating images within a video as a 'bag'.
• Introduced a co-person attention mechanism to leverage shared identities across videos.

Main Results:

• The attention weights are derived from all images, making the model robust to noisy annotations.
• Achieved superior performance compared to existing methods on two weakly labeled person re-identification datasets.

Conclusions:

• The proposed weakly supervised method effectively performs person re-identification using only video-level labels.
• The attention-based framework offers a scalable and robust solution for person re-identification in videos.