A Complete Content-based 3D Shape Retrieval System Using Deep Learning

This work addresses the problem of content-based 3D shape retrieval. Despite the variety of approaches proposed in the literature, the challenge that still lies ahead is to design a method that allows large-scale retrieval and respects relevance. In this paper, we offer a complete system that takes full advantage of deep learning. Using Convolutional Neural Network (CNN) for shape indexing, our key idea is to consider the final output (softmax layer) of a well-trained model as a given 3D object’s descriptor. This solution is employed not only to improve relevance but also to accelerate the shape matching process. In addition to serving as the search key, the considered descriptor is used to predict the models` list to match. Hence, this will avoid a sequential and systematic comparison with all the 3D models in the database. To evaluate our approach with state-of-the-art methods, we use the SHREC`17 event specifications and ShapeNetCore subset data. We demonstrate that the proposed shape indexing technique improves the relevance and accelerates the retrieval process significantly.