Stereo Vision for Driving: Unsupervised Learning of Depth Perception

Stacey Svetlichnaya

Like most deep learning problems, computer vision for autonomous driving is best solved by adding more labeled data. This is often prohibitively slow and expensive. In a CVPR 2017 paper, Tinghui Zhou’s team presents an unsupervised framework for estimating depth and motion from monocular video, such as a car dashboard camera. They train two networks in parallel: one to predict depth from a single frame and the other to predict the current view from several frames (e.g. the previous and next frames). At testing time, they only use the first network, enabling depth perception from a single photo. Dive into examples and details with the full report=>

Join our mailing list to get the latest machine learning updates.