This repository contains an open source implementation of a CVPR 2018 paper which reconstructs the depth of a scene based on a color image of it.

The paper achieves competitive results by:

• reusing an existing CNN architecture (ResNet-152) and extending it with some new, specialized convolutional layers

• introducing a new loss function which improves the accuracy of the depth estimation

• running the network on the image and on smaller cropped sections of the image, then recombining the results using Fourier domain analysis

The original implementation, available on the authors' site, was written in MATLAB and used Caffe. This implementation is based on the PyTorch library for doing deep learning in Python.

The code in this repository is licensed under the Mozilla Public License Version 2.0, see the LICENSE file for the full text.