tinghuiz/SfMLearner

Uses a photometric loss on consecutive video frames to jointly train depth and pose estimation networks end-to-end, eliminating the need for ground-truth annotations. Built on TensorFlow 1.0 and supports training on KITTI and Cityscapes datasets with evaluation tools provided for standard benchmarks. The architecture learns geometric constraints from frame synthesis during video sequences rather than supervised depth labels.

2,014 stars. No commits in the last 6 months.