senguptaumd / Background-Matting

Background Matting: The World is Your Green Screen
https://grail.cs.washington.edu/projects/background-matting/
4.78k stars 663 forks source link

Background Matting: The World is Your Green Screen

alt text

By Soumyadip Sengupta, Vivek Jayaram, Brian Curless, Steve Seitz, and Ira Kemelmacher-Shlizerman

This paper will be presented in IEEE CVPR 2020.

Project Page

Go to Project page for additional details and results.

Paper (Arxiv)

Blog Post

Background Matting v2.0

We recently released a brand new background matting project: better quality and REAL-TIME performance (30fps at 4K and 60fps at FHD)! You can now use this with Zoom! Much better quality! We tried this on a Linux machine with a GPU.

Check out the code!

Project members

Acknowledgement: Andrey Ryabtsev, University of Washington

License

This work is licensed under the Creative Commons Attribution NonCommercial ShareAlike 4.0 License.

Summary

Updates

April 21, 2020:

April 20,2020

April 9, 2020

April 8, 2020

Getting Started

Clone repository:

git clone https://github.com/senguptaumd/Background-Matting.git

Please use Python 3. Create an Anaconda environment and install the dependencies. Our code is tested with Pytorch=1.1.0, Tensorflow=1.14 with cuda10.0

conda create --name back-matting python=3.6
conda activate back-matting

Make sure CUDA 10.0 is your default cuda. If your CUDA 10.0 is installed in /usr/local/cuda-10.0, apply

export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64
export PATH=$PATH:/usr/local/cuda-10.0/bin

Install PyTorch, Tensorflow (needed for segmentation) and dependencies

conda install pytorch=1.1.0 torchvision cudatoolkit=10.0 -c pytorch
pip install tensorflow-gpu==1.14.0
pip install -r requirements.txt

Note: The code is likely to work on other PyTorch and Tensorflow versions compatible with your system CUDA. If you already have a working environment with PyTorch and Tensorflow, only install dependencies with pip install -r requirements.txt. If our code fails due to different versions, then you need to install specific CUDA, PyTorch and Tensorflow versions.

Run the inference code on sample images

Data

To perform Background Matting based green-screening, you need to capture:

Use sample_data/ folder for testing and prepare your own data based on that. This data was collected with a hand-held camera.

Pre-trained model

Please download the pre-trained models from Google Drive and place Models/ folder inside Background-Matting/.

Note: syn-comp-adobe-trainset model was trained on the training set of the Adobe dataset. This was the model used for numerical evaluation on Adobe dataset.

Pre-processing

  1. Segmentation

Background Matting needs a segmentation mask for the subject. We use tensorflow version of Deeplabv3+.

cd Background-Matting/
git clone https://github.com/tensorflow/models.git
cd models/research/
export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
cd ../..
python test_segmentation_deeplab.py -i sample_data/input

You can replace Deeplabv3+ with any segmentation network of your choice. Save the segmentation results with extension _masksDL.png.

  1. Alignment

Skip this step, if your data is captured with fixed-camera.

Run python test_pre_process.py -i sample_data/input for pre-processing. It aligns the background image _back.png and changes its bias-gain to match the input image _img.png

We used AKAZE features python code (since SURF and SIFT unavilable in opencv3) for alignment. We also provide an alternate MATLAB code (test_pre_process.m), which uses SURF features. MATLAB code also provides a way to visualize feature matching and alignment. Bad alignment will produce bad matting output. Bias-gain adjustment is turned off in the Python code due to a bug, but it is present in MATLAB code. If there are significant exposure changes between the captured image and the captured background, use bias-gain adjustment to account for that.

Feel free to write your own alignment code, choose your favorite feature detector, feature matching and alignment.

Background Matting

python test_background-matting_image.py -m real-hand-held -i sample_data/input/ -o sample_data/output/ -tb sample_data/background/0001.png

For images taken with fixed camera (with a tripod), choose -m real-fixed-cam for best results. -m syn-comp-adobe lets you use the model trained on synthetic-composite Adobe dataset, without real data (worse performance).

Run the inference code on sample videos

This is almost exactly similar as that of the image with few small changes.

Data

To perform Background Matting based green-screening, you need to capture:

We provide sample_video/ captured with hand-held camera and sample_video_fixed/ captured with fixed camera for testing. Please download the data and place both folders under Background-Matting. Prepare your own data based on that.

Pre-processing

  1. Frame extraction:
cd Background-Matting/sample_video
mkdir input background
ffmpeg -i teaser.mov input/%04d_img.png -hide_banner
ffmpeg -i target_back.mov background/%04d.png -hide_banner

Repeat the same for sample_video_fixed

  1. Segmentation
    cd Background-Matting/models/research/
    export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
    cd ../..
    python test_segmentation_deeplab.py -i sample_video/input

Repeat the same for sample_video_fixed

  1. Alignment

No need to run alignment for sample_video_fixed or videos captured with fixed-camera.

Run python test_pre_process_video.py -i sample_video/input -v_name sample_video/teaser_back.png for pre-processing. Alternately you can also use test_pre_process_video.m in MATLAB.

Background Matting

For hand-held videos, like sample_video:

python test_background-matting_image.py -m real-hand-held -i sample_video/input/ -o sample_video/output/ -tb sample_video/background/

For fixed-camera videos, like sample_video_fixed:

python test_background-matting_image.py -m real-fixed-cam -i sample_video_fixed/input/ -o sample_video_fixed/output/ -tb sample_video_fixed/background/ -b sample_video_fixed/teaser_back.png

To obtain the video from the output frames, run:

cd Background-Matting/sample_video
ffmpeg -r 60 -f image2 -i output/%04d_matte.png -vcodec libx264 -crf 15 -s 1280x720 -pix_fmt yuv420p teaser_matte.mp4
ffmpeg -r 60 -f image2 -i output/%04d_compose.png -vcodec libx264 -crf 15 -s 1280x720 -pix_fmt yuv420p teaser_compose.mp4

Repeat same for sample_video_fixed

Notes on capturing images

For best results capture images following these guidelines:

Training on synthetic-composite Adobe dataset

Data

Training

Change number of GPU and required batch-size, depending on your platform. We trained the model with 512x512 input (-res flag).

CUDA_VISIBLE_DEVICES=0,1 python train_adobe.py -n Adobe_train -bs 4 -res 512

Notes:

Cheers to the unofficial Deep Image Matting repo.

Training on unlabeled real videos

Data

Please download our captured videos.. We will show next how to finetune your model on fixed-camera captured videos. It will be similar for hand-held cameras, except you will need to align the captured background image to each frame of the video separately. (Take a hint from test_pre_process.py and use alignImages().)

Data Pre-processing:

Training

Change number of GPU and required batch-size, depending on your platform. We trained the model with 512x512 input (-res flag).

CUDA_VISIBLE_DEVICES=0,1 python train_real_fixed.py -n Real_fixed -bs 4 -res 512 -init_model Models/syn-comp-adobe-trainset/net_epoch_64.pth

Dataset

We captured videos with both fixed and hand-held camera in indoor and outdoor settings. We release this data to encourage future research on improving background matting. The data is released for research purposes only.

Download data

Google Colab

Thanks to Andrey Ryabstev for creating Google Colab version for easy inference on images and videos of your choice.

Google Colab

Notes

We are eager to hear how our algorithm works on your images/videos. If the algorithm fails on your data, please feel free to share it with us at soumya91@cs.washington.edu. This will help us in improving our algorithm for future research. Also, feel free to share any cool results.

Citation

If you use this code for your research, please consider citing:

@InProceedings{BMSengupta20,
  title={Background Matting: The World is Your Green Screen},
  author = {Soumyadip Sengupta and Vivek Jayaram and Brian Curless and Steve Seitz and Ira Kemelmacher-Shlizerman},
  booktitle={Computer Vision and Pattern Regognition (CVPR)},
  year={2020}
}

Related Implementations

Microsoft Virtual Stage: Using our background matting technology along with depth sensing with Kinect, Microsoft opensourced this amazing code for virtual staging. Follow this link for details of their technique.

Weights & Biases: Great presentation and detailed discussions and insights on pre-processing and training our model. Check out Two Minutes Paper's take on our work.