DiffusionDC: Depth Completion Using Deformable Diffusion Model

Zehang Shen, Jiuming Liu, Haidong Hu, Yong Wang, Fan Xu

Abstract: Image-guided depth completion is a challenging task that aims to predict dense depth maps from RGB images and sparse depth maps. Previous methods mainly rely on affinity-based approaches, such as spatial propagation networks (SPNs). However, the rigid refinement process in SPNs highly depends on the feature extraction capabilities of the backbone network and fails to explicitly model noise. In this work, we propose DiffusionDC, a novel depth completion network based on the diffusion probabilistic model. The iterative denoising refinement layer is designed to correct the errors in coarse depth maps using visual features. To better utilize neighboring information, we incorporate deformable convolution layers into the denoising network. Our DiffusionDC achieves rather competitive performance on the NYUv2 indoor dataset and KITTI outdoor dataset.

External IDs:dblp:conf/rcar/ShenLHWX25