Object Bounding Box-Aware Embedding for Point Cloud Instance Segmentation

In 2D image domain, recent researches have made significant progress in encoding context information for instance segmentation. While the counterpart in point cloud is still left far behind. Previous works mostly focus on leveraging semantic information and aggregating point local information through K-Nearest-Neighbor method. Such methods are unaware of object boundary information which is important to separating nearby objects. We propose a novel module to integrate object bounding box information into embedding for Point Cloud Instance Segmentation. The proposed module called Object Bounding Box-aware module (OBAM) boosts the instance segmentation performance by encoding Object Bounding Box information. Through attention mechanism, the module removes redundant boundary information. Comprehensive experiments on two popular benchmarks (S3DIS and ScanNetV2) show the effectiveness of our method. Our method achieves the State-of-the-art instance segmentation performance on S3DIS benchmark.