Squeezing bag-of-features for scalable and semantic 3D model retrieval

We have previously proposed a multiple-view, densely-sampled, bag-of-visual features algorithm for shape-based 3D model retrieval. The method achieved good retrieval performance for moderately sized benchmark datasets (~1,000 3D models), including both rigid and articulated 3D shapes. It is also much faster than the other methods having similar retrieval performance. However, the method does not exploit semantic knowledge. We want the retrieval results to reflect multiple (e.g., ~100) semantic classes. Also, if applied to a larger database (e.g., ~1M models), search through the database can be expensive due to its large feature vector size (e.g., 30k dimensions). This paper proposes a method to "squeeze" its length feature vector by projecting it onto a manifold that incorporates multiple semantic classes. Experimental evaluation has shown that retrieval performances equal or better than original feature can be achieved while reducing feature vector size, e.g., from 30k down to less than 100.