PIIE-DSA-Net for 3D Semantic Segmentation of Urban Indoor and Outdoor Datasets

Abstract

In this paper, a 3D semantic segmentation method is proposed, in which a novel feature extraction framework is introduced assembling point initial information embedding (PIIE) and dynamic self-attention (DSA)—named PIIE-DSA-net. Ideal segmentation accuracy is a challenging task, since the sparse, irregular and disordered structure of point cloud. Currently, taking into account both low-level features and deep features of the point cloud is the more reliable and widely used feature extraction method. Since the asymmetry between the length of the low-level features and deep features, most methods cannot reliably extract and fuse the features as expected and obtain ideal segmentation results. Our PIIE-DSA-net first introduced the PIIE module to maintain the low-level initial point-cloud position and RGB information (optional), and we combined them with deep features extracted by the PAConv backbone. Secondly, we proposed a DSA module by using a learnable weight transformation tensor to transform the combined PIIE features and following a self-attention structure. In this way, we obtain optimized fused low-level and deep features, which is more efficient for segmentation. Experiments show that our PIIE-DSA-net is ranked at least in the top seventh among the most recent published state-of-art methods on the indoor dataset and also made a great improvement than original PAConv on outdoor datasets.