Direct Joint Detection from Humanoid 3D Models without using Skeleton Information.

Skeletonization, or automatic skeleton extraction, is one of the most essential technologies in 3DCG. This technology makes it possible to automatically extract skeletons (i.e. bones, joints and their hierarchical structures) from 3D models. Such skeletons are important shape and pose descriptors for object representation, object recognition etc. They are used in many applications such as 3D model search, virtual characterʹs pose estimation and collision detection between two or more 3D models. However, existing skeletonization methods have some drawbacks. Most of the existing skeletonization methods have difficulties in correctly extracting the positions of joints. In most methods, bones are extracted from a 3D model first and joints are defined as the cross points of bones. However some errors always occur when bones are extracted. Hence joints cannot be found in this scheme so often. Furthermore, they are not allowing for controlling the number of bones/joints and their structure. Therefore applying motion data acquired from motion capture devices to 3D models still involves a lot of cumbersome manual work. In this paper, we propose a novel joint detection method suited for kinematic skeleton generation, skeleton rigging etc. Unlike the existing methods, the proposed method detects joint positions directly without using skeleton (bone) information. So the proposed method can avoid propagating errors occurred by skeletonization process. Also, the proposed method is able to extract the same numbers of joints/bones and the same structure as in given motion data, i.e. one can directly apply existing motion data without the need of manual adjustment. In general, 3D models describe shape information and pose information simultaneously. Distinguishing one from the other seems to be very difficult. However, the proposed method solves this problem by extracting only the pose information of 3D models by using a vertex Gauss sphere representation and estimating the positions of joints correctly regardless of shapes of 3D models by adopting template matching approach. Experimental results showed that the proposed method achieves 90 % accuracy of pose estimation and 73% accuracy of joint estimation.)