DetPoseNet: Improving Multi-Person Pose Estimation via Coarse-Pose Filtering

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human detection and pose estimation are crucial for analyzing human activities in visual data.
• Current top-down methods struggle with occlusions and crowded scenes due to early detection commitments.
• Failures in initial human detection lead to inaccuracies in subsequent pose estimation.

Purpose of the Study:

• To introduce DetPoseNet, a novel end-to-end framework for unified multi-human detection and pose estimation.
• To overcome limitations of traditional top-down approaches in complex scenarios.
• To enhance the accuracy and efficiency of human pose estimation.

Main Methods:

• DetPoseNet employs a unified three-stage network: coarse-pose proposal extraction, proposal filtering, and multi-scale pose refinement.
• The framework generates initial whole-body bounding boxes and keypoint proposals in a single shot.
• Multi-scale supervision, multi-scale regression, structure-aware loss, and keypoint masking are utilized for robust pose refinement.

Main Results:

• The proposed framework effectively filters unlikely detections, improving subsequent pose estimation accuracy.
• Experiments on COCO and OCHuman datasets validate the effectiveness of DetPoseNet.
• The method achieves a significant computational efficiency, with a 5-6x speedup in multi-person pose estimation.

Conclusions:

• DetPoseNet provides a robust and efficient solution for multi-human detection and pose estimation.
• The unified framework addresses challenges posed by occlusions and crowded scenes.
• The approach demonstrates superior performance and speed compared to existing methods.