Hamdi Abdullah ; Giancola Silvio ; Ghanem Bernard

Abstract

Multi-view projection methods have demonstrated their ability to reachstate-of-the-art performance on 3D shape recognition. Those methods learndifferent ways to aggregate information from multiple views. However, thecamera view-points for those views tend to be heuristically set and fixed forall shapes. To circumvent the lack of dynamism of current multi-view methods,we propose to learn those view-points. In particular, we introduce theMulti-View Transformation Network (MVTN) that regresses optimal view-points for3D shape recognition, building upon advances in differentiable rendering. As aresult, MVTN can be trained end-to-end along with any multi-view network for 3Dshape classification. We integrate MVTN in a novel adaptive multi-view pipelinethat can render either 3D meshes or point clouds. MVTN exhibits clearperformance gains in the tasks of 3D shape classification and 3D shaperetrieval without the need for extra training supervision. In these tasks, MVTNachieves state-of-the-art performance on ModelNet40, ShapeNet Core55, and themost recent and realistic ScanObjectNN dataset (up to 6% improvement).Interestingly, we also show that MVTN can provide network robustness againstrotation and occlusion in the 3D domain. The code is available athttps://github.com/ajhamdi/MVTN .