AutoPedestrian: An Automatic Data Augmentation and Loss Function Search Scheme for Pedestrian Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Pedestrian detection is crucial but challenging, especially in crowded environments with frequent occlusions.
• Existing methods struggle with the complexities of dense pedestrian scenarios.

Purpose of the Study:

• To propose AutoPedestrian, a novel scheme for enhancing pedestrian detection performance.
• To automatically optimize data augmentation strategies and loss functions jointly for improved accuracy in crowded scenes.

Main Methods:

• Formulating data augmentation and loss functions as probability distributions.
• Employing a double-loop scheme with importance sampling for efficient joint optimization.
• Developing an automated approach for selecting optimal augmentation policies and loss functions.

Main Results:

• Achieved state-of-the-art results on CrowdHuman and CityPersons benchmarks.
• Demonstrated significant improvements in pedestrian detection accuracy, particularly in challenging crowded scenarios.
• Reported 40.58% MR on CrowdHuman and 11.3% MR on the CityPersons reasonable subset.

Conclusions:

• The proposed AutoPedestrian scheme effectively enhances pedestrian detection in crowded scenes.
• Jointly searching for optimal data augmentation and loss functions is a promising direction for computer vision tasks.
• The method sets new benchmarks for performance in pedestrian detection research.