Pose Mask: A Model-Based Augmentation Method for 2D Pose Estimation in Classroom Scenes Using Surveillance Images.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning for pose estimation shows promise but struggles with dense crowds and occlusion.
• Existing image augmentation methods lack specific knowledge for pose estimation challenges.

Purpose of the Study:

• To develop an effective pose estimation method for occluded classroom scenes.
• To leverage masked autoencoder (MAE) capabilities for image augmentation in pose estimation.

Main Methods:

• Proposed a top-down pose estimation approach utilizing MAE's reconstruction ability.
• Introduced 'Pose Mask,' an augmentation technique referencing keypoint distribution heatmaps instead of random masking.
• Collected a new dataset, Class Pose, specifically for classroom pose estimation.

Main Results:

• The proposed method demonstrated promising performance in heavily occluded classroom scenes.
• Pose Mask augmentation effectively addressed occlusion challenges in pose estimation.
• Pre-trained MAE weights showed utility as a model-based augmentation strategy.

Conclusions:

• The MAE-based approach with Pose Mask is effective for occluded classroom pose estimation.
• This method offers a novel solution for improving pose estimation accuracy in challenging real-world scenarios.
• The Class Pose dataset facilitates further research in this area.