Sketch-based 3D Shape Retrieval with Multi-Silhouette View Based on Convolutional Neural Networks

Abstract

Searching for relevant 3D models based on hand-drawn sketches is important for many applications, such as sketch-based 3D modeling and recognition. Sketch-based shape retrieval (SBSR) has become a hot research spot in the field of model retrieval, pattern recognition, and computer vision. 3D deep representation based on Convolutional Neural Network (CNN) enables significant performance improvement over state-of-the-arts in task of 3D shape retrieval. Motivated by this, we proposed a sketch-based 3D model retrieval algorithm by utilizing representative views and CNN feature matching. The representative views are obtained by viewpoint entropy. The main idea of the method is that the hand-drawn sketch can be achieved according to a viewpoint of the 3D model. Thus the sketch and the projection of model from same class are similar. Therefore, we filter a certain amount of view as representative view to reduce the computational complexity and improve the accuracy. We extract CNN descriptors as features for representative view of each object. Our experiments on Shape Retrieval Contest (SHREC) 2012 database and SHREC 2013 database demonstrate that our method is better than state-of-the-art approaches.