High-Resolution Self-attention with Fair Loss for Point Cloud Segmentation

Applying deep learning techniques to analyze point cloud data has emerged as a prominent research direction. However, the insufficient spatial and feature information integration within point cloud and unbalanced classes in real-world datasets have hindered the advancement of research. Given the success of self-attention mechanisms in numerous domains, we apply the High-Resolution Self-Attention (HRSA) module as a plug-and-play solution for point cloud segmentation. The proposed HRSA module preserve high-resolution internal representations in both spatial and feature dimensions. Additionally, by affecting the gradient of dominant and weak classes, we introduce the Fair Loss to address the problem of unbalanced class distribution on a real-world dataset to improve the network’s inference capabilities. The introduced modules are seamlessly integrated into an MLP-based architecture tailored for large-scale point cloud processing, resulting in a new segmentation network called PointHR. PointHR achieves impressive performance with mIoU scores of 69.8% and 74.5% on S3DIS Area-5 and 6-fold cross-validation. With a significantly smaller number of parameters, these performances make PointHR highly competitive in point cloud semantic segmentation.