wpfhtl / 3net

Repository for "Learning monocular depth estimation with unsupervised trinocular assumptions"

3net

This repository contains the source code of 3net, proposed in the paper "Learning monocular depth estimation with unsupervised trinocular assumptions", 3DV 2018. If you use this code in your projects, please cite our paper:

@inproceedings{3net18,
  title     = {Learning monocular depth estimation with unsupervised trinocular assumptions},
  author    = {Matteo Poggi and
               Fabio Tosi and
               Stefano Mattoccia},
  booktitle = {6th International Conference on 3D Vision (3DV)},
  year = {2018}
}

For more details: arXiv

Demo video: youtube

Requirements

• Tensorflow 1.8 (recomended)
• python packages such as opencv, matplotlib

Run 3net on webcam stream

To run 3net, just launch

sh get_checkpoint.sh
python webcam.py --checkpoint_dir /checkpoint/3DV18/3net --mode [0,1,2]

While the demo is running, you can press:

• 'm' to change mode (0: depth-from-mono, 1: depth + view synthesis, 2: depth + view + SGM)
• 'p' to pause the stream
• 'ESC' to quit

Train 3net from scratch

Code for training will be (eventually) uploaded. Meanwhile, you can train 3net by embedding it into https://github.com/mrharicot/monodepth