SfMLearner implements depth and ego-motion estimation from monocular videos using self-supervised learning. The system trains deep learning models to jointly predict scene depth and camera motion by leveraging view synthesis as the supervisory signal. It was published at CVPR 2017 and supports training on KITTI and Cityscapes datasets. The codebase requires TensorFlow 1.0 and CUDA, and includes pre-trained models for single-view depth prediction demos.