self-supervised-depth-completion

Why we included this project

This is the reference PyTorch implementation of the Sparse-to-Dense method from ICRA 2019, which fuses sparse LiDAR depth samples with a monocular RGB image to produce dense depth maps. Developed at MIT, it was trained on the KITTI dataset alone, with no pretraining on Cityscapes or other driving data. The repo covers both training regimes: supervised training against semi-dense LiDAR ground truth, and a self-supervised mode that relies on photometric, sparse-depth, and smoothness losses, so you can compare the two directly. It also includes KITTI download scripts, pretrained checkpoints, and command-line options for training, resuming, and evaluating. If you need to densify sparse range measurements for driving perception or robotics, this is a solid reference to start from, though it is research code rather than a maintained production service.