Dense Depth Posterior (DDP) from Single Image and Sparse Range
This addresses depth completion for autonomous driving and robotics, but it is incremental as it builds on existing methods with improvements.
The paper tackles the problem of estimating dense depth maps from single images and sparse range measurements, achieving state-of-the-art results on the KITTI benchmark for both unsupervised and supervised depth completion.
We present a deep learning system to infer the posterior distribution of a dense depth map associated with an image, by exploiting sparse range measurements, for instance from a lidar. While the lidar may provide a depth value for a small percentage of the pixels, we exploit regularities reflected in the training set to complete the map so as to have a probability over depth for each pixel in the image. We exploit a Conditional Prior Network, that allows associating a probability to each depth value given an image, and combine it with a likelihood term that uses the sparse measurements. Optionally we can also exploit the availability of stereo during training, but in any case only require a single image and a sparse point cloud at run-time. We test our approach on both unsupervised and supervised depth completion using the KITTI benchmark, and improve the state-of-the-art in both.