Multi-stage feature point detection for 3D human data

• Authors: Xiang Pan;Alexander Agathos
• Venue: J-HGBU '11 Proceedings of the 2011 joint ACM workshop on Human gesture and behavior understanding
• Year: 2011

In this paper, we present an automatic approach to detect feature points on 3D human models. Instead of simultaneously detecting all feature points, as previous approaches do, our algorithm recursively detect feature points by using a multi-stage strategy. The multi-stage strategy firstly detects the extreme feature points giving them also semantic labels such as head, hands and feet using human domain knowledge. Then the detected points and their semantic labels are used to find other feature points such as crotch, arms, knees and armpits. Finally, a graph cut is used to refine the position of the detected feature points. Besides its robustness under different poses, our algorithm has also two distinct advantages: Full automation and high efficiency. In the experiments made, we verify the proposed algorithm by using a plethora of human models with different poses and size.