TR-Net: A Transformer-Based Neural Network for Point Cloud Processing

Point cloud is a versatile geometric representation that could be applied in computer vision tasks. On account of the disorder point cloud, it challenging to design deep neural network used analysis. Furthermore, most existing frameworks for processing either hardly consider local neighboring information or ignore context-aware and spatially-aware features. To deal with above problems, we propose novel architecture named TR-Net, which based on transformer. This reformulates task as set-to-set translation problem. TR-Net directly operates raw clouds without any data transformation annotation, reduces consumption computing resources memory usage. Firstly, neighborhood embedding backbone designed effectively extract from cloud. Then, an attention-based sub-network constructed better learn semantically abundant discriminatory embedded Finally, effective global features are yielded through feeding extracted by into residual backbone. For different downstream tasks, build decoders. Extensive experiments public datasets illustrate our approach outperforms other state-of-the-art methods. example, performs 93.1% overall accuracy ModelNet40 dataset archives mIou 85.3% ShapeNet part segmentation.