DT-Net: Point Cloud Completion Network With Neighboring Adaptive Denoiser and Splitting-Based Upsampling Transformer

Point cloud completion, which involves inferring missing regions of 3D objects from partial observations, remains a challenging problem in 3D vision and robotics. Existing learning-based frameworks typically leverage an encoder-decoder architecture to predict the complete point cloud based on the global shape representation extracted from the incomplete input, or further introduce a refinement network to optimize the obtained complete point cloud in a coarse-to-fine manner, which is unable to capture fine-grained local geometric details and filled with noisy points in the thin or complex structure. In this article, we propose a novel coarse-to-fine point cloud completion framework called DT-Net, by focusing on coarse point cloud denoising and multi-level upsampling. Specifically, we propose a Neighboring Adaptive Denoiser (NAD) to effectively denoise the coarse point cloud generated by an autoencoder, and reduce noise around the slender structures, making them clear and well represented. Moreover, a novel Splitting-based Upsampling Transformer (SUT), which effectively incorporates spatial and semantic relationships between local neighborhoods in the point cloud, is also proposed for multi-level upsampling. Extensive qualitative and quantitative experiments demonstrate that our method outperforms state-of-the-art methods under widely used benchmarks.