Bag-of-Features Sampling Techniques for 3D CAD Model Retrieval

This paper investigates two sampling strategies, dense sampling and PHOW sampling, for bag-of-features 3D CAD model retrieval. Previous methods [1] use original salient SIFT feature detection for general 3D model retrieval which does not suit the need for CAD models representation. CAD models contain mostly piecewise-smooth surfaces and thus only sharp edges can be described. To overcome these limitations, two new sampling strategies are investigated to improve the feature extraction process. Dense sampling extracts SIFT features on regular spatial grids with even spacing. Pyramid Histogram Of visual Words (PHOW) [2] extracts features on repeatedly finer scales. Both the two sampling methods extract features that are covered the whole shape. In addition, the influences of codebook size and distance metric are also studied to optimize the retrieval performance. Experiments on Purdue Engineering Benchmark [3] show that the proposed sampling techniques achieve better retrieval accuracy than the original salient SIFT sampling method.