Adaptive Graph Convolution Algorithm Based on 3D Vision Selectivity and Its Application in Scene Segmentation
Ning Ma, Songwen Jin, Yanming Zhao · Traitement du signal · 2024
At this stage, the 3D graph convolution algorithm has the following problems: (1) Neighbor space selection problem; (2) Feature extraction and fusion problem of different depth map convolution algorithms; (3) Multi-view parallel feature fusion problem.Based on this, "Multi-domain adaptive graph convolution algorithm based on visual computing theory and its scene segmentation application" is proposed.First, inspired by the 3D vision of primates, a 3D visual computing theory is proposed; and propose an adaptive graph convolution algorithm based on the 3D visual selectivity theory.It solves the problem of neighbor space selection for 3D graph convolution; secondly, inspired by the single-link serial processing mode of primate visual information, a single-link depth adaptive graph convolution algorithm based on 3D visual selectivity is constructed to learn and refuse the different depth visual features of the same sub-space of 3D point cloud; Finally, inspired by the multi-link parallel processing model of primate visual information, we improved the single-link algorithm and constructed a multi-link depth adaptive graph convolution algorithm based on 3D visual selectivity to learn and integrate global visual features of different link; and using the MLP algorithm with shared weights to achieve object segmentation.On ShapeNetPart and custom Mortise_and_Tenon_DB, Compare with PointNet, PointNet++, KPConv deform, 3D GCN and other algorithms.Verify the segmentation performance and geometric invariance of this algorithm.The experimental results: The segmentation performance of the algorithm of this article is good, and the segmentation success rate reaches 90.9%; the algorithm of this article has strong geometric invariance, Rotation and translation transformations are geometrically invariant, and Scaling transformation has finite geometric invariance in the interval [-0.15,0.15].