Point Cloud Networks with Adaptive Sampling and Cross-Domain Feature Interaction

As an efficient characterization paradigm for 3D scene understanding, point clouds have been widely used in computer vision and robot perception. However, its unstructured nature is susceptible to noise interference, occlusion, density and scale variation problems, and there are challenges such as feature loss and boundary geometry distortion. To this end, this study proposes ASCFNet, a hierarchical point cloud network based on adaptive sampling and cross-domain feature interaction. the multi-scale adaptive sampling module (MAS) implements a dynamic sampling strategy that fuses local features with global information through the Gumbel-softmax function. Cross-domain feature interaction module (CFI), which employs a dual mechanism of intersection fusion and local context aggregation to enhance the learning ability of boundary features. Symmetric Feature Propagation Architecture (SFPA), which effectively suppresses feature bias in the traditional up-sampling process through consistency constraints in the encoding-decoding path. Experiments on benchmark datasets such as ModelNet40 and S3DIS show that the method achieves an average accuracy of 92.5 % in the classification task (1.1 % improvement over the benchmark model) and 74.4% mIoU in the semantic segmentation task, which validates the advantages of the framework.