Point Cloud Upsampling via Cascaded Refinement Network

Point cloud upsampling focuses on generating a dense, uniform and proximity-to-surface point set. Most previous approaches accomplish these objectives by carefully designing single-stage network, which makes it still challenging to generate high-fidelity distribution. Instead, in coarse-to-fine manner is decent solution. However, existing methods require extra training strategies, are complicated time-consuming during the training. In this paper, we propose simple yet effective cascaded refinement consisting of three generation stages that have same network architecture but achieve different objectives. Specifically, first two dense coarse points progressively, while last stage further adjust better position. To mitigate learning conflicts between multiple decrease difficulty regressing new points, encourage each predict offsets with respect input shape. manner, proposed can be easily optimized without strategies. Moreover, design transformer-based feature extraction module learn informative global local shape context. inference phase, dynamically model efficiency effectiveness, depending available computational resources. Extensive experiments both synthetic real-scanned datasets demonstrate approach outperforms state-of-the-art methods. The code publicly at https://github.com/hikvision-research/3DVision .