Multi-scale Feature Cross Perception Instance Segmentation Network based on 3D Point Cloud

Abstract

To address challenges in 3D point cloud instance segmentation, we propose a novel network called Multi-scale Feature Cross Perception Instance Segmentation Network (MFCPNet). MFCPNet incorporates two key enhancements to improve the performance of 3D point cloud instance segmentation. Firstly, we introduce a multi-scale semantic aggregation module that effectively captures scene features using Markov Chain Monte Carlo (MCMC) for point filtering and aggregation. This module enables the extraction of intricate semantic information across various scales. Secondly, MFCPNet integrates a feature cross perception module employing an attention mechanism to independently process features from different scales. By connecting features from diverse scales, the model achieves a comprehensive and enriched representation. Extensive experiments on the ScanNetv2 dataset demonstrate the superiority of MFCPNet. Notably, MFCPNet outperforms other methods by 2.8% in mean Average Precision (mAP) on the ScanNetv2 test set.