PointNet++SAKS: A Point Cloud Model Based on KANs and Attention Mechanism for Objects Classification and Semantic Segmentation

Abstract

Traditional point clouds models based on multilayer perceptrons (MLPs) lack inherent support for spatial data structures and fail to effectively process spatial relationships. To address this limitation, we propose PointNet++ SAKS, a deep learning network for point clouds processing. The network incorporates a local feature aggregation module that replaces the structure of MLPs with Kolmogorov-Arnold networks (KANs), and is able to capture complex higher-order spatial dependencies in point clouds through its recursive and adaptive network structure. Unlike MLPs, which rely only on layer-by-layer weighted linear transformations, KANs introduce more nonlinear and hierarchical interactions in modeling the local structure, making it better able to understand and model the complex relationships between points in a local region. The introduced SegNext combines Transformer and Convolutional Neural Network (CNN) to extract features from different scales using convolutional kernels of different sizes or convolutional operations with different step sizes. On the Sydney Urban Objects Dataset classification, overall accuracy (OA) and mean accuracy (mAcc) metrics improved by 4.89% and 9.47%, respectively. On ModelNet40, OA and mAcc are improved by 4.9% and 5.2%, respectively, and training time is reduced. For the semantic segmentation task, the OA, mAcc, and intersection over union (IoU) metrics are improved by 1.1%, 4.6%, and 3.7%, respectively. We show that PointNet++SAKS outperforms the baseline model with high accuracy and speed.

Keywords

• PointNet++
• deep learning
• point clouds classification
• attention mechanism
• semantic segmentation