Depth estimation is a process of computing 3rd dimension from and 2D image which was lost during image acquisition phase. Depth estimation plays an important role to understand the geometry of the scene. There are various approaches to compute depth estimation from an image Supervised, Unsupervised and Semi-Supervised approached. In this work, unsupervised approach from the literature is investigated to train ML model using KITTI driving dataset, following the research paper Unsupervised Monocular Depth Estimation with Left-Right Consistency. Most of the techniques relay on Supervised training approached which require ground truth data in addition to images data. To cope with this challenge, unsupervised learning scheme is investigated which does not require ground truth data with images. Instead, model trained in unsupervised fashion on binocular images will learn disparity map between left and right images, during training cycles. The trained model learns the image disparity map as image reconstruction problem for depth estimations from the monocular vision.
SlimLoigx 2020. All Rights Reserved