Awesome

SSN: Soft Shadow Network for Image Composition (CVPR 2021 Oral)

Results

Preparation

Remember to recursively git clone this repository.

git clone --recurse-submodules -j8 https://github.com/ShengCN/SSN_SoftShadowNet.git

Data Render

The renderer for the data is in Data folder. This renderer is accelerated by CUDA. The computation time will be proportional to final image dimension and triangle numbers.

Environment Requirements

System: Ubuntu 18.04
CMake: >= 3.8
GCC: >= 7.4.0
CUDA (Ensure CMake can find nvcc)

CMake is used to manage the packages and compiling. All other dependencies required except CUDA are in the data/src/Dep folders. Again, remember to recursively clone the submodules.

There is a pre-compiled Ubuntu(18.04) version in data/src/build folder. If you cannot run the data/src/build/run.sh script, then you need to recompile.

(Optional) Recompile

cd data/src/build
rm -r CMake*
cmake ..
make -j16

Command Interface

Change directory to Data/src
Put obj/off files to models folder
Check the parameters in the render.sh, e.g. gpus(support multiple gpu), random pitch range, random model rotation?

# File render.sh
model_root="models"
cache="Dataset/human_data/cache"
start_id=0
end_id=-1 # -1 means all models
width=512
height=512
cpus=6
out_hdf5="Dataset/human_data/all_base.hdf5"
cam_pitch_min=0
cam_pitch_max=45
model_rot_min=-90
model_rot_max=90
samples=2

rm tmp/scene_cache.bin

python render.py --cpus=$cpus --gpus 0 2 --model_root=$model_root --out_folder=$cache --start_id=$start_id --end_id=$end_id --width=$width --height=$height --samples=$samples --cam_pitch_min=$cam_pitch_min --cam_pitch_max=$cam_pitch_max --model_rot_min=$model_rot_min --model_rot_max=$model_rot_max &&
python build_hdf5.py --cache=$cache --width=$width --height=$height --out_hdf5=$out_hdf5

Run ./render.sh The output will be a hdf5 file, which composed of all the data. The file structure for this hdf5:

x/scene_name          # Input for training    
rgb/scene_name        # Rendered RGB for testing
base_32_8/scene_name  # Shadow base in 32 x 8 IBL resolution. 
base_64_16/scene_name # Shadow base in 64 x 16 IBL resolution

To see how to use this hdf5 file, please refer to Train/datasets/SSN_Dataset.py.

Training

We prepared a training script. After rendering the dataset, please modify the hdf5_file entry in Train/configs/ssn.yaml accordingly. Run Train/train.sh will begin the training for SSN.

Bonus

I implemented a general training framework in the Train/ folder. It is YAML driven. The minimal work to use this framework for your experiment is to write a dataloader and define the model architectures.

About

License

SSN may be used non-commercially, meaning for research or evaluation purposes only.

Citation

@inproceedings{sheng2021ssn,
  title={SSN: Soft shadow network for image compositing},
  author={Sheng, Yichen and Zhang, Jianming and Benes, Bedrich},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4380--4390},
  year={2021}
}