CAS-SR: convolutional additive self-attention for range-view-based point cloud super-resolution

Compared to directly acquiring data using high-precision LiDAR devices, point cloud super-resolution offers a more cost-effective alternative for obtaining high-accuracy 3D spatial information. It enhances the spatial resolution of point clouds and improves performance in downstream tasks. To reduce training resource consumption and accelerate inference, many existing approaches attempt to convert the 3D point cloud super-resolution task into a 2D image super-resolution problem. However, although such methods improve computational efficiency to some extent, they often rely on complex and time-consuming attention mechanisms for feature extraction, which still require substantial computational resources. To address this issue, this paper proposes a novel point cloud super-resolution network architecture, CAS-SR, designed to achieve fast and high-quality point cloud up-sampling with reduced computational cost. The proposed method consists of three main stages: projecting sparse point clouds into 2D image space, extracting key features from the images, and generating a high-resolution range image, which is then back-projected into 3D space for super-resolution reconstruction. Experimental results demonstrate that CAS-SR maintains high reconstruction accuracy while significantly improving inference speed, showing promising real-time performance and strong potential for practical deployment.