Automatic extraction of generic focal features on 3D shapes via random forest regression analysis of geodesics-in-heat

Flexible definition and automatic extraction of generic features on 3D shapes is important for feature-centric geometric analysis, however, existing techniques fall short in measuring and locating semantic features from users' psychological standpoints. This paper makes an initial attempt to propose a learning based generic modeling approach for user-central definition and automatic extraction of features on arbitrary shapes. Instead of purely resorting to certain local geometric extremes to simply formulate feature metrics, it enables the users to arbitrarily specify application-specific features on training shapes, so that similar features can be automatically extracted from other same-category shapes with isometric or near-isometric deformations. Our key originality is built upon an observation: the geodesic distance from one point to desired feature point on testing shape should be similar to the points on training shapes. To this end, we propose a novel regression model to bridge the massive random-sampled local properties and the desired feature points via incorporating their corresponding geodesic distances into the powerful random forest framework. On that basis, an effective voting strategy is proposed to estimate the locations of the user-specified features on new shapes. Our extensive experiments and comprehensive evaluations have demonstrated many attractive advantages of our method, including being fully-automatic, robust to noise and partial holes, invariant under isometric and near-isometric deformation, and also scale-invariant, which can well facilitate to many downstream geometry-processing applications such as semantic mesh segmentation, mesh skeleton extraction, etc.