A Self-Training Approach for Point-Supervised Object Detection and Counting in Crowds.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection in crowded scenes is challenging due to occlusions and high object density.
• Existing methods often require dense annotations, which are labor-intensive to acquire.

Purpose of the Study:

• To develop a novel self-training approach for object detection using only point-level annotations.
• To enable accurate estimation of both center points and sizes of crowded objects.

Main Methods:

• Crowd-SDNet utilizes point annotations for direct center point estimation.
• It employs a crowdedness-aware loss and a confidence/order-aware refinement scheme for object size regression.
• An effective decoding method is proposed for extremely crowded scenes.

Main Results:

• Crowd-SDNet significantly outperforms state-of-the-art point-supervised methods on the WiderFace benchmark.
• Achieved over 10% improvement in average precision and a 31.2% reduction in counting error.
• Achieved state-of-the-art results on crowd counting, localization, and vehicle counting datasets.

Conclusions:

• The proposed self-training approach effectively handles object detection in crowded environments using minimal annotations.
• Crowd-SDNet demonstrates superior performance in both detection and counting tasks, offering a practical solution for real-world applications.