Foreground-Aware Geometry Compression With Hybrid Attention for Large-Scale Point Clouds

Regions of Interest (ROI) play a crucial role in point cloud compression, especially in applications such as autonomous driving and robot navigation, where foreground regions often contain key information such as obstacles and object boundaries. However, traditional point cloud compression methods typically fail to optimize for these critical areas, instead applying uniform processing across the whole point clouds. The paper aims to optimize the point cloud compression process by allocating more bitstream resources to the foreground regions, thereby preserving important information in the point clouds. To achieve this, we propose a separation-based Foreground-Background Network (FB-Net) for compressing point cloud. The framework first identifies and separates the foreground and background regions, then designs an attention-based compression network, which includes multi-stage Occupancy Probability Estimation (OPE) module. The OPE module consist of an Attention-based Feature Extraction Layer (AFEL) and an Occupancy Probability Generation (AOPG) module. By controlling the number of OPE modules, we can allocate more bitstream resources to critical regions in the scene point cloud, such as tables and chairs, thus improving the performance of downstream detection tasks. Furthermore, to compensate for perceptual distortions in human vision, we design a large-scale receptive field-based Point Cloud Upsampling Network (PCU-Net), to enhance the objective quality. Through extensive experiments on point cloud datasets such as ScanNet and SUN RGB-D, we demonstrate that allocating more bitstream resources to the foreground regions benefits the accuracy of detection tasks. Compared to G-PCC and many state-of-the-art learning-based point cloud compression methods, our approach shows superior performance in detection tasks than compression-then-detection process methods.