Learning diffusion on global graph: A PDE-directed approach for feature detection on geometric shapes

Feature and saliency analyses are crucial for various graphics applications. The key idea is to automatically compute and recommend the salient or outstanding regions of concerned models. However, there is no universally-applicable criterion for the detection results stemming from the personalized viewpoints for interest features on each specific model. This paper proposes a human-oriented feature detection framework, learning diffusion on global graph (LDGG), to understand personalized interests in a simple and low-cost way. A user-friendly interaction method is introduced to incorporate specific human interests as detection criteria in a small training set. Given a test model, we model the interest feature detection process as partial differential equations (PDEs)-directed diffusion on the global graph composed of nodes extracted from all training and test models. To infer the real interest points of users, submodular optimization is employed to select the source seeds adaptively for the diffusion system. By introducing diffusion guidance based on interest information, the PDEs become learnable. Extensive experiments and comprehensive comparisons have exhibited many attractive advantages of the proposed framework, such as capable of small-sample learning, easy-to-implement, extendable, self-correction, discriminative power, etc.