SfMLearner

Why we included this project

This is the authors' reference implementation of a 2017 CVPR paper on learning depth and camera motion from monocular video without ground-truth depth labels. The network works by reconstructing consecutive frames from its predicted depth and ego-motion, with photometric consistency as the training signal, a trick that later driving and robotics systems built on. The repo includes training and evaluation code for KITTI and Cityscapes, a pre-trained single-view depth model, and an interactive demo notebook, so you can reproduce the paper's results or adapt the architecture to your own footage. Keep in mind the code targets TensorFlow 1.0-era tooling, so treat it as a research baseline and learning reference rather than a drop-in production pipeline. For teams starting out in self-supervised monocular depth estimation, it's a clear, well-documented place to begin.