AdaBins

Why we included this project

AdaBins is the official implementation of the CVPR 2021 paper on adaptive binning for monocular depth estimation, and it remains a solid reference if you need per-pixel depth from a single RGB image. Instead of dividing the depth range into fixed intervals, the network learns where to place its bins from the image content, which is what gives it better accuracy on both indoor and outdoor scenes. The repo includes pretrained weights for NYU-Depth-v2 and KITTI, plus an InferenceHelper class that handles preprocessing and test-time augmentation, so you can move from a downloaded checkpoint to predicted depth maps without much setup. It is a clean baseline to adapt for depth work in robotics or AR, though the project has not been updated since 2022.