Efficient Person Search via Expert-Guided Knowledge Distillation.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Person search in images is computationally demanding, requiring both high accuracy and efficiency for practical use.
• Existing methods often struggle to balance performance with speed.

Purpose of the Study:

• To develop a unified framework for person search that effectively addresses both accuracy and efficiency demands.
• To leverage knowledge distillation (KD) for creating compact and fast person search models.

Main Methods:

• Proposed a unified person search framework integrating an efficient student network with deep expert networks.
• Employed knowledge distillation (KD) in a multitask learning approach to transfer expertise to a compact model.
• Evaluated three customized student networks of varying scales.

Main Results:

• Achieved strong performance on par with state-of-the-art methods in terms of mean average precision and top-1 accuracies.
• Demonstrated high efficiency with a feedforward time of 120 frames/second.
• Showcased minimal accuracy sacrifice despite significant efficiency gains.

Conclusions:

• The proposed unified framework effectively balances accuracy and efficiency in person search.
• Knowledge distillation is a viable technique for developing compact and high-performing person search models.
• The method shows promise for real-world person search applications requiring speed and precision.