What it does This repo compiles ranked results for depth completion—turning sparse depth maps (from LiDAR or visual odometry) into dense ones, guided by RGB images. It tracks both supervised and unsupervised/self-supervised methods across two benchmarks: VOID (indoor, from visual-inertial odometry) and KITTI (outdoor, from LiDAR).

The interesting bit The author ranks by all four metrics (MAE, RMSE, iMAE, iRMSE) rather than letting papers cherry-pick RMSE. The tables also link directly to code, so you can verify whether the numbers hold up.

Key highlights

• Covers 9+ unsupervised and 7+ supervised methods on VOID; 11+ unsupervised and 10+ supervised on KITTI
• Includes 2024 papers (AugUndo, OGNI-DC, BP-Net) alongside older baselines like IP Basic’s classical CPU method
• Most entries have linked PyTorch or TensorFlow implementations
• Maintained by Alex Wong, who also authored several of the top-ranked methods

Caveats

• README is truncated mid-table in the supervised KITTI section; full rankings aren’t visible
• No explanation of what iMAE/iRMSE mean or why they matter for ranking
• Update cadence is unclear—some 2024 papers are in, but coverage may lag

Verdict Worth bookmarking if you’re picking a depth completion method for robotics or 3D vision. Skip it if you need training code or tutorials; this is pure benchmark curation.