3D point cloud has tremendous potential in many application tasks. However, the huge amount of data limits this potential. To simplify point clouds and improve their downstream application efficiency, this paper proposes AS-Net, an attention-aware downsampling network oriented to classification tasks. AS-Net realizes downsampling through an Attention-aware Sampling Module, which including an Input Embedding Module and an Attention Module. The former is designed to extract the global and local features of the point cloud, the latter is to generate the Sampling-Map to simulate the differentiable downsampling. Thanks to the attention mechanism, AS-Net may select the critical points of the original point cloud for classification tasks. In addition, AS-Net designs a Constraint Matching Module to match the sampled points to be a subset of the original point cloud at the inference phase. For end-to-end training, AS-Net construct a joint loss function that includes a task loss, a sampling loss, and a constraint loss. Extensive experiments on the ModelNet10/40 and ShapeNet datasets demonstrate that AS-Net achieves a good performance on the point cloud classification task. Especially when the downsampling size is small, the result is better than the referenced methods.
展开▼
