Multi-scale single-stage pose detection with adaptive sample training in the classroom scene

Abstract

Student pose information plays an important role in teaching management and evaluation. Thus, it is meaningful to acquire student pose information fast and precisely. Although the pose estimation techniques are well-studied recently, it is still challenging to apply existing methods to handle this task due to heavy occlusion in the educational environment. Different from the pose estimation, we consider pose detection as a specific problem of object detection. However, compared with generic object detection problems, pose detection usually suffers from poor detection performance due to the challenges of similar categories, object with small sizes, quantity imbalance of categories, etc. To address these issues, we propose a new pose detection method based on a single-stage object detector. We present a multi-scale feature enrichment branch to obtain balanced and robust features. Then we adopt an adaptive fusion mechanism to learn complementary spatial features, making our feature extractor more discriminative. Besides, an adaptive positive sample training strategy is adopted to select robust positive samples and make full use of high-quality predicted positive samples when training by the adaptive Smooth L1 loss. Experimental results show that the proposed method obviously outperforms other single-stage object detection methods on the real classroom pose datasets.