关键词: 草图检索, 三维模型, 词袋模型, 流形排序

Abstract: Currently, retrieve 3d models use hand-drawn sketch as query has become an active research field. In the Bag of Feature (BoF) framework, the granularity of the distinction and expression brought by the "hard clustering" is insufficient. To solve this problem, a new method was proposed: GALIF feature was extracted from each sketch and the set of view images of models. BOF was replaced by a feature coding based on “soft clustering”--Gaussian mixture model. A fast and effective manifold ranking algorithm was used to calculate the manifold similarity between images, giving the final retrieve results. The previous methods only considered the data similarity, like directly calculated Euclidean distance between images. Different from those methods, the method tried to capture the semantic relevance by the underlying structure of the data set. The experiment result demonstrates that the new method is more accurate than the state-of-arts methods.

Key words: sketch retrieve, 3d models, fisher coding, manifold ranking