Category-specific upright orientation estimation for 3D model classification and retrieval

Abstract

In this paper, we address a problem of correcting upright orientation of a reconstructed object to search. We first reconstruct an input object appearing in an image sequence, and generate a query shape using multi-view object co-segmentation. In the next phase, we utilize the Convolutional Neural Network (CNN) architecture to determine category-specific upright orientation of the queried shape for 3D model classification and retrieval. As a practical application of our system, a shape style and a pose from an inferred category and up-vector are obtained by comparing 3D shape similarity with candidate 3D models and aligning its projections with a set of 2D co-segmentation masks. We quantitatively and qualitatively evaluate the presented system with more than 720 upfront-aligned 3D models and five sets of multi-view image sequences.