View-relation constrained global representation learning for multi-view-based 3D object recognition

Multi-view observations provide complementary clues for 3D object recognition, but also include redundant information that appears different across views due to view-dependent projection, light reflection and self-occlusions. This paper presents a view-relation constrained global representation network (VCGR-Net) for 3D object recognition that can mitigate the view interference problem at all phases, from view-level source feature generation to multi-view feature aggregation. Specifically, we determine inter-view relations via LSTM implicitly. Based on the relations, we construct a two-stage feature selection module to filter features at each view according to their importance to the global representation and their reliability as observations at specific views. The selected features are then aggregated by referring to intra- and inter-view spatial context to generate global representation for 3D object recognition. Experiments on the ModelNet40 and ModelNet10 datasets demonstrate that the proposed method can suppress view interference and therefore outperform state-of-the-art methods in 3D object recognition.