Deformable convolutional networks for multi-view 3D shape classification

This Letter suggests a novel method for improving the robustness and the geometric transformation modelling capability in multi-view convolutional networks (MVCNNs). First, the deformable convolutional networks are used to learn more details and features related to the geometric transformation which the standard convolutional neural networks cannot handle. Then a view-pooling layer is specifically designed for combining the descriptors from multiple views as the final representations of the 3D shapes. The key idea is to insert the deformable convolutional layer between the input and convolutional layer, making it possible to solve deformable 3D shape classification problems, which was a challenging task for MVCNN framework. The proposed method achieves state-of-the-art classification results on two subsets of the ModelNet dataset (ModelNet10 and ModelNet40) over previous methods by a significant margin.