Single-Image Depth Perception in the Wild

Code for reproducing the results in the following paper:

Single-Image Depth Perception in the Wild,
Weifeng Chen, Zhao Fu, Dawei Yang, Jia Deng
Neural Information Processing Systems (NIPS), 2016.

Please check out the project site for more details.

Setup

1. Install the Torch 7 framework as described in http://torch.ch/docs/getting-started.html#_. Please make sure that you have the cudnn, hdf5 and csvigo modules installed.

2. Clone this repo.

    git clone https://github.com/wfchen-umich/relative_depth.git
   

3. Download and extract the DIW dataset from the project site. Download and extract DIW_test.tar.gz and DIW_train_val.tar.gz into 2 folders. Run the following command to download and extract DIW_Annotations.tar.gz. Then modify the filepath to images in DIW_test.csv, DIW_train.csv and DIW_val.csv to be the absolute file path where you extracted DIW_test.tar.gz and DIW_train_val.tar.gz.

    cd relative_depth
    mkdir data
    cd data
    wget https://vl-lab.eecs.umich.edu/data/nips2016/DIW_Annotations_splitted.tar.gz
    tar -xzf DIW_Annotations_splitted.tar.gz
    rm DIW_Annotations_splitted.tar.gz
   

Training and evaluating the networks

Testing on pre-trained models

Please first run the following commands to download the test data from our processed NYU dataset and the pre-trained models:

cd relative_depth
wget https://vl-lab.eecs.umich.edu/data/nips2016/data.tar.gz
tar -xzf data.tar.gz
rm data.tar.gz
cd data
python convert_csv_2_h5.py -i 750_train_from_795_NYU_MITpaper_train_imgs_800_points_resize_240_320.csv
python convert_csv_2_h5.py -i 45_validate_from_795_NYU_MITpaper_train_imgs_800_points_resize_240_320.csv

cd ../src
mkdir results
cd results
wget https://vl-lab.eecs.umich.edu/data/nips2016/hourglass3.tar.gz
tar -xzf hourglass3.tar.gz
rm hourglass3.tar.gz

Note: You can also download data.tar.gz and hourglass3.tar.gz from Google Drive.

Then change directory into /relative_depth/src/experiment.

1. To evaluate the pre-trained model Ours(model trained on the NYU labeled training subset) on the NYU dataset, run the following command:

    th test_model_on_NYU.lua -num_iter 1000 -prev_model_file ../results/hourglass3/NYU_795_800_c9_1e-3/Best_model_period1.t7 -test_set 654_NYU_MITpaper_test_imgs_orig_size_points.csv -mode test -thresh 0.9
   

2. To evaluate the pre-trained model Ours_Full(model trained on the full NYU training set) on the NYU dataset, run the following command:

    th test_model_on_NYU.lua -num_iter 1000 -prev_model_file ../results/hourglass3/1e-3_Drop_205315_NYU_fs_c9/Best_model_period1.t7 -test_set 654_NYU_MITpaper_test_imgs_orig_size_points.csv -mode test -thresh 0.32
   

3. To evaluate the pre-trained model Ours_DIW(our network trained from scratch on DIW) on the DIW dataset, run the following script:

    th test_model_on_DIW.lua -num_iter 90000 -prev_model_file ../results/hourglass3/AMT_from_scratch_1e-4_release/Best_model_period1.t7 -test_model our
   

4. To evaluate the trained model Ours_NYU_DIW(our network pre-trained on NYU and fine-tuned on DIW) on the DIW dataset, run the following script:

    th test_model_on_DIW.lua -num_iter 90000 -prev_model_file ../results/hourglass3/AMT_from_205315_1e-4_release/Best_model_period2.t7 -test_model our
   

5. To test on a single image, we provide a handy script test_on_one_image.lua:

    th test_on_one_image.lua -prev_model_file Model.t7 -input_image input.jpg -output_image output.jpg
   

Training

Please first change directory into /relative_depth/src/experiment.

To train the model Ours(model trained on the NYU labeled training subset), please run the following command:

th main.lua -lr 0.001 -bs 4 -m hourglass3 -it 100000 -t_depth_file 750_train_from_795_NYU_MITpaper_train_imgs_800_points_resize_240_320.csv -v_depth_file 45_validate_from_795_NYU_MITpaper_train_imgs_800_points_resize_240_320.csv -rundir ../results/hourglass3/Ours

To train the model Ours_DIW(our network trained from scratch on DIW), please run the following command:

th main.lua -diw -lr 0.000100 -bs 4 -m hourglass3 -it 200000 -t_depth_file DIW_train.csv -v_depth_file DIW_val.csv -rundir ../results/hourglass3/Ours_DIW

To train the model Ours_NYU_DIW(our network pre-trained on NYU and fine-tuned on DIW), please run the following command:

cd relative_depth/src/results/hourglass3/
mkdir Ours_NYU_DIW
cp 1e-3_Drop_205315_NYU_fs_c9/Best_model_period1.t7 Ours_NYU_DIW/205315_Best_model_period1.t7
cd ../../experiment/
th main.lua -diw -lr 0.000100 -bs 4 -m hourglass3 -it 200000 -t_depth_file DIW_train.csv -v_depth_file DIW_val.csv -start_from 205315_Best_model_period1.t7 -rundir ../results/hourglass3/Ours_NYU_DIW/

More

This work is also featured in the first release of the Wolfram Neural Net Repository. See this article for more details.