3D model retrieval based on multi-view attentional convolutional neural network

摘要

We propose a discriminative Multi-View Attentional Convolutional Neural Network, dubbed as MVA-CNN, which takes the multiple views of an shape as input and output the object category. Unlike previous view-based approaches that simply "compile" the view features into a compact 3D descriptors, our method can discover the context among multiple views in both the visual and spatial domain. First, we extract multiple rendered images from a 3D object by virtual cameras, and then we use Convolutional Neural Network (CNN) to abstract the information of the views. Second, we aggregate the visual views by two steps: 1). an element-wise maximum operation across the view features is adopted to discover discriminative features. 2). a soft attention mechanism is used to dynamically adjust the shape descriptors for better representing the spatial information. The entire network can be trained in an end-to-end way with the standard backpropagation. We verify the effectiveness of MVA-CNN on two widely used datasets: ModelNetlO, ModelNet40 by comparing our method with state-of-the-art methods.