Estimating a right-view from a monocular image(to make a stereo pair of images) which respects the geometry of the scene is an important problem in computer vision. This repository aims to achieve this by implementing Deep3D (see original repo) using PyTorch to generate right view of images. This generated pair of images can then be used to estimate depth in images, convert 2D video to 3D, etc.
We used the KITTI Stereo 2015 Dataset. The dataset consists of 200 training scenes and 200 test scenes, which include 4 color images per scene, in a lossless png format. This means that we have 400 left and right image pairs for training.
The pretrained model can be downloaded from here
The following are some results of our approach (from left to right- input left , ground truth right stereo image pair, generated stereo image pair, depth map generated using OpenCV for image pairs):
