Code and data for paper "Shading Annotations in the Wild".

Our code was tested on Ubuntu 14.04. As a first step, clone our repo:

git clone https://github.com/kovibalu/saw_release.git

Then install the python dependencies by running:

sudo ./install/install_python.sh

If you would like to run our trained model, you will need to install Caffe. We slightly modified the implementation of Bansal et. al for our purposes. To check out our Caffe version which is included as a submodule, run:

git submodule update --init --recursive

Then build Caffe after editing the Makefile.config depending on your configuration with:

cd caffe
make all -j
make pycaffe -j

To download all data related to the dataset, run:

./download_saw.sh

The whole dataset download size is ~28.0GB, please see the documentation in the script for a detailed breakdown of sizes for the different parts of the dataset. For detailed documentation on the format of the downloaded annotations in saw/saw_annotations_json see ANNO_FORMAT.md.

To generate the precision-recall curves in our paper for all baselines and our method, run:

python main.py generate_pr

You can select which baselines to evaluate in main.py.

To generate the pixel labels from the SAW annotations and NYUv2 depth dataset depth and normal maps, run:

python main.py generate_labels

Please cite our paper if you use our code or data:

@article{kovacs17shading,
	author = "Balazs Kovacs and Sean Bell and Noah Snavely and Kavita Bala",
	title = "Shading Annotations in the Wild",
	journal = "Computer Vision and Pattern Recognition (CVPR)",
	year = "2017",
}

Please contact Balazs Kovacs with any questions.