taherahmadi / Leveraging-Adversarial-training-for-Monocular-Depth-Estimation

adding a discriminator network and introducing two new loss functions for monocular depth estimation

Leveraging-Adversarial-training-for-Monocular-Depth-Estimation

Achieving higher accurcay in the details of the objects in depth maps by adding a discriminator network, adversarial training, and introducing two new loss functions for monocular depth estimation

Parham Yassini*, Taher Ahmadi*, Elnaz Mehrzadeh*, Dorsa Dadjoo*, Fatemeh Hasiri*

*Equal Contribution

Results

Dependencies

• python 3.7
  
• Pytorch 1.3.1
  

Running

Download the trained models and put in the root of project: Depth estimation networks
Download the data and put it in the the following structure: NYU-v2 dataset
.(project root)/data/
├── nyu2_test
├── nyu2_test.csv
├── nyu2_train
└── nyu2_train.csv

• Demo
  

  python demo.py
  

• Test
  

  python test.py
  

• Training
  

  python train.py
  

Citation

this work is a extension on the: Revisiting Single Image Depth Estimation: Toward Higher Resolution Maps With Accurate Object Boundaries. Junjie Hu et al.