Highly Efficient Active Learning With Tracklet-Aware Co-Cooperative Annotators for Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Supervised person re-identification (ReID) shows great potential but is limited by high human annotation costs.
• Annotating identical pedestrians across different cameras is expensive and time-consuming.

Purpose of the Study:

• To propose a novel framework that significantly reduces annotation demand in person ReID.
• To maintain high performance in person ReID while minimizing the need for human input.

Main Methods:

• A tracklet-aware co-cooperative annotators' framework is developed.
• Training samples are clustered, and adjacent images form robust tracklets to decrease annotation needs.
• An active learning strategy with a teacher model selects informative tracklets for human annotation, while the teacher also labels certain tracklets.

Main Results:

• The proposed framework achieves competitive performance on popular person ReID datasets.
• The approach demonstrates effectiveness in both active learning and unsupervised learning (USL) settings.
• The method successfully trains models using both pseudo-labels and human annotations.

Conclusions:

• The tracklet-aware co-cooperative annotators' framework effectively reduces annotation costs in person ReID.
• This approach offers a viable solution for large-scale person ReID applications with limited annotation budgets.
• The integration of active learning and pseudo-labeling enhances model training efficiency and performance.