Pedestrian Detection by Exemplar-Guided Contrastive Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Existing pedestrian detection methods struggle with occlusions, scale variations, and diverse appearances (silhouettes, viewpoints, clothing).
• Individual learning of diverse pedestrian features is a significant challenge in current approaches.

Purpose of the Study:

• To develop a novel contrastive learning framework for pedestrian detection that addresses appearance diversity.
• To minimize semantic distance between pedestrians with varying appearances and maximize distance from background elements.

Main Methods:

• Utilized contrastive learning to guide feature learning, reducing appearance diversity.
• Constructed an exemplar dictionary of representative pedestrian appearances to create effective contrastive training pairs.
• Employed the exemplar dictionary to evaluate pedestrian proposal quality during inference.

Main Results:

• The proposed method effectively minimizes appearance diversities among pedestrians in the learned feature space.
• Maximized the distance between pedestrians and background, enhancing detection accuracy.
• Demonstrated effectiveness through extensive experiments on both daytime and nighttime pedestrian detection datasets.

Conclusions:

• The contrastive learning approach with an exemplar dictionary significantly improves pedestrian detection, especially under challenging appearance variations.
• The method offers a robust solution for real-world pedestrian detection scenarios, including varying lighting and viewpoints.