DcTr: Noise-robust point cloud completion by dual-channel transformer with cross-attention

Ben Fei, Weidong Yang, Lipeng Ma, Wenming Chen

Abstract: Highlights•Our DcTr focus on the explicitly and locally structured pattern in the points generation process, achieving superior results to existing methods.•Our DcTr proposes a dual-channel transformer to learn the points generation patterns in the decoder.•DcTr integrates cross-attention to incorporate the local region information from the encoder into the points features of the decoder.•DcTr has excellent robustness against different amounts and intensifiers of noise, proving the usefulness in the real world.