Unordered images selection for dense 3D reconstruction based on Distance Dependent Chinese Restaurant Process

Abstract

Key frame selection is important to dense 3D reconstruction, especially for unordered image sets. A novel method for key frame selection from unordered image sets is proposed based Distance Depedent Chinese Restaurant Process (DDCRP). First, a bag-of-features word package is constructed to describe each image in a document-like manner, which can be dealt with by the DDCRP model. Second, the overlapping measure among images is computed by the idea borrowed from image stitching framework. Third, all images are clustered into different clusters by the DDCRP according to their bag-of-feature description and overlapping measure. Finally, a subset of images is chosen from all clusters and used for the 3D construction. To verify the effectiveness of the proposed method, we tested it based on a number of image collections. Experimental results show that the proposed method can effectively remove most of the redundant images, and consequently reduce time cost and improve robustness for dense 3D reconstruction.