This repository is the release of MSG-Net for our paper Depth Map Super-Resolution by Deep Multi-Scale Guidance in ECCV16. It comes with four trained networks (x2, x4, x8, and x16), one hole-filled RGBD training set, and three hole-filled RGBD testing sets (A, B, and C).
To the best of our knowledge, MSG-Net is the FIRST convolution neural network which attempts to 1) upsample depth images and 2) use the multi-scale guidance from the corresponding high-resolution RGB images.
Another repository for MS-Net (without multi-scale guidance) is also available.
For more details, please visit the project page .
We train our models using caffe and evaluate the results on Matlab.
You need to install caffe and remeber to complie matcaffe. You can put the folder MSGNet-release in caffe/examples. Finally, you need to get into the the directory of examples/MSGNet-release/util, and run MSGNet.m.
Our RGBD training set consists of 58 RGBD images from MPI Sintel depth dataset, and 34 RGBD images from Middlebury dataset. 82 images are used for training and 10 images (frames 1, 20, 28, 58, 64, 66, 69, 73, 75 and 79) are used for validation.
Testig set is available /MSGNet-release/testing sets.
All code is provided for research purposes only and without any warranty. Any commercial use requires our consent. If our work helps your research or you use the code in your research, please cite the following paper:
@InProceedings{hui2016,
author = {Tak-Wai Hui and Chen Change Loy and and Xiaoou Tang},
title = {Depth Map Super-Resolution by Deep Multi-Scale Guidance},
booktitle = {Proceedings of European Conference on Computer Vision (ECCV)},
pages = {353--369},
year = {2016},
url = {http://mmlab.ie.cuhk.edu.hk/projects/guidance_SR_depth.html}
}