“Revisiting Deformable Convolution for Depth Completion” is a research project that proposes an innovative approach to depth completion, a task which aims to produce high-quality dense depth maps from sparse ones. In previous methods, refining depth maps often requires multiple iterations and suffers from limitations due to a fixed receptive field. The researchers in this study address these challenges by leveraging deformable convolution, leading to a single-pass refinement module. The model is evaluated on the KITTI dataset, demonstrating top-tier performance in terms of both accuracy and inference speed.
Publication date: 3 Aug 2023
Project Page: https://github.com/AlexSunNik/ReDC
Paper: https://arxiv.org/pdf/2308.01905.pdf
