Sparse-Depth-Completion

Why we included this project

Autonomous driving perception has to turn sparse, irregular LiDAR returns into dense depth maps, and this is the reference implementation of a method that does exactly that. It fuses monocular RGB images with the LiDAR points, splitting the job into a global and a local branch whose predictions are merged using confidence maps the model learns without supervision. At the time the paper appeared, this approach topped the KITTI depth completion benchmark, and a strong baseline like that is worth keeping around: trained weights and test predictions are published, so teams benchmarking their own depth models can reproduce the result directly. Just note the license. It allows personal and research use only, so treat the code as a solid comparison point rather than something to fold into a commercial product.