Few-Shot Depth Completion Using Denoising Diffusion Probabilistic Model

Abstract

Generating dense depth maps from sparse LiDAR data is a challenging task, benefiting a lot of computer vision and photogrammetry tasks including autonomous driving, 3D point cloud generation, and aerial spatial awareness. Using RGB images as guidance to generate pixel-wise depth map is good, but these multi-modal data fusion networks always need numerous high-quality datasets like KITTI dataset to train on. Since this may be difficult in some cases, how to achieve few-shot learning with less train samples is worth discussing. So in this paper, we firstly proposed a few-shot learning paradigm for depth completion based on pre-trained denoising diffusion probabilistic model. To evaluate our model and other baselines, we constructed a smaller train set with only 12.5% samples from KITTI depth completion dataset to test their few-shot learning ability. Our model achieved the best on all metrics with a 5% improvement in RMSE compared to the second-place model.