Ranking on semantic manifold for shape-based 3d model retrieval

Semantics associated with 3D shapes are often as important as the shapes themselves in defining "shape similarity" among them. So far, only a small subset of 3D model retrieval methods took semantics into account. Most popular approach to semantic 3D model retrieval is based on Relevance Feedback (RF), an iterative, interactive approach for a system to learn a semantic class that embodies "user intention" for the query. A drawback of a typical RF-based method is its low initial performance as it starts cold without any semantic knowledge. An alternative approach is off-line learning of multiple semantic classes. The approach produces a good retrieval performance without per-query training iterations, but is unable to capture user intention per-query. The method proposed in this paper attempts to combine benefits of the two approaches so that both shared multiple semantic classes and per-query intention can be captured to improve 3D model retrieval. Our method first learns, off-line, the multiple semantic classes by using a semi-supervised manifold learning algorithm to produce a "semantic manifold" of the input features. The RF iteration based on manifold ranking algorithm is then run on the semantic manifold. Our empirical evaluation showed that this method significantly outperforms the manifold ranking run in the original, ambient feature space.