nianticlabs/monodepth2

Self-supervised learning framework that trains on monocular or stereo video sequences without ground-truth depth labels, using photometric loss and view synthesis. Implements ResNet encoder-decoder architecture with multi-scale depth predictions and supports multiple training modalities (mono, stereo, mono+stereo) on KITTI dataset. Built in PyTorch with pretrained models available at various resolutions (640×192 to 1024×320) and extensible dataloader API for custom datasets.

4,466 stars. No commits in the last 6 months.