Code for reproducing the results in the following paper:

Learning Single-Image Depth from Videos using Quality Assessment Networks
Weifeng Chen, Shengyi Qian, Jia Deng
Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

Please check the project site for more details.

1. The code is written in python 2.7.13, using pytorch 0.2.0_4. Please make sure that you install the correct pytorch version as later versions may cause the code to break.

2. Clone this repo.

git clone [email protected]:princeton-vl/YouTube3D.git

3. Download data_model.tar.gz into path YouTube3D, then untar:

cd YouTube3D
tar -xzvf data_model.tar.gz

4. Download and unpack the images from Depth in the Wild dataset. Edit DIW_test.csv under YouTube3D/data so that all the image paths are absolute paths.

To evaluate the pre-trained model EncDecResNet trained on ImageNet + ReDWeb + DIW + YouTube3D on the DIW dataset, run the following command:

cd YouTube3D/src 
python test.py -t DIW_test.csv -model exp/YTmixReD_dadlr1e-4_DIW_ReDWebNet_1e-6_bs4/models/model_iter_753000.bin

In case you want to get the qualitative outputs, append a -vis flag and the qualitative outputs will be in the folder visualize:

mkdir visualize
python test.py -t DIW_test.csv -model exp/YTmixReD_dadlr1e-4_DIW_ReDWebNet_1e-6_bs4/models/model_iter_753000.bin -vis

To evaluate the pre-trained model HourglassNetwork trained on NYU + DIW + YouTube3D on the DIW dataset, run the following command:

python test.py -t DIW_test.csv -model exp/Hourglass/models/best_model_iter_852000.bin

Please send any questions or comments to Weifeng Chen at [email protected].