This repository contains an implementation of the video object segmentation method FRTM. A detailed description of the method is found in the CVPR 2020 paper "Learning Fast and Robust Target Models for Video Object Segmentation"

CVF: [paper] [supplement]
Arxiv: [paper]

If you find the code useful, please cite using:

@InProceedings{Robinson_2020_CVPR,
    author = {Robinson, Andreas and Lawin, Felix Jaremo and Danelljan, Martin and Khan, Fahad Shahbaz and Felsberg, Michael},
    title = {Learning Fast and Robust Target Models for Video Object Segmentation},
    booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2020}

}

Clone the repository: git clone https://github.com/andr345/frtm-vos.git

Create a conda environment and install the following dependencies:

sudo apt install ninja-build  # For Debian/Ubuntu
conda install -y cython pip scipy scikit-image tqdm
conda install -y pytorch torchvision cudatoolkit=10.1 -c pytorch
pip install opencv-python easydict

PyTorch 1.0.1 is slightly faster. If you wish to try to this version, replace the conda install pytorch above with the following:

conda install pytorch==1.0.1 torchvision==0.2.2 -c pytorch
pip install "pillow<7"

To test the DAVIS validation split, download and unzip the 2017 480p trainval images and annotations here: https://davischallenge.org/davis2017/code.html.

Or, more precisely, this file.

To test our validation split and the YouTubeVOS challenge 'valid' split, download YouTubeVOS 2018 and place it in this directory structure:

/path/to/ytvos2018
|-- train/
|-- train_all_frames/
|-- valid/
`-- valid_all_frames/

You only actually need 300 sequences of train/ and train_all_frames/ and these are listed in lib/ytvos_jjvalid.txt. Thanks to Joakim Johnander for providing this split.

These pretrained models are available for download:

The script weights/download_weights.sh will download all models and put them in the folder weights/.

Open evaluate.py and adjust the paths dict to your dataset locations and where you want the output. The dictionary is found near line 110, and looks approximately like this:

    paths = dict(
        models=Path(__file__).parent / "weights",  # The .pth files should be here
        davis="/path/to/DAVIS",  # DAVIS dataset root
        yt2018="/path/to/ytvos2018",  # YouTubeVOS 2018 root
        output="/path/to/results",  # Output path
    )

Then try one of the evaluations below. The first run will pause for a few seconds while compiling a PyTorch C++ extension.

Scripts for generating the results in the paper:

python evaluate.py --model rn101_ytvos.pth --dset yt2018val       # Ours YouTubeVos 2018
python evaluate.py --model rn101_all.pth --dset dv2016val         # Ours DAVIS 2016
python evaluate.py --model rn101_all.pth --dset dv2017val         # Ours DAVIS 2017

python evaluate.py --model rn18_ytvos.pth --fast --dset yt2018val # Ours fast YouTubeVos 2018
python evaluate.py --model rn18_all.pth --fast --dset dv2016val   # Ours fast DAVIS 2016
python evaluate.py --model rn18_all.pth --fast --dset dv2017val   # Ours fast DAVIS 2017

--model is the name of the checkpoint to use in the weights directory.

--fast reduces the number of optimizer iterations to match "Ours fast" in the paper.

--dset is one of

Training is set up similarly to evaluation.

Open train.py and adjust the paths dict to your dataset locations, checkpoint and tensorboard output directories and the place to cache target model weights.

To train a network, run

python train.py <session-name> --ftext resnet101 --dset all --dev cuda:0

--ftext is the name of the feature extractor, either resnet18 or resnet101.

--dset is one of dv2017, ytvos2018 or all ("all" really means "both").

--dev is the name of the device to train on.

Replace "session-name" with whatever you like. Subdirectories with this name will be created under your checkpoint and tensorboard paths.

Training target models from scratch and filling the cache take approximately 5 days of training. Once the cache is mostly full, the next training session should take less than 24 hours. The cache requires 17 GB disk space for training with ResNet-101 features and 32 intermediate channels (as in the paper) and 5 GB for ResNet-18 and the same number of channels.

Our own cache (20 GB) is available here. The link is not permanent and will change eventually, so make sure to check this readme in the GitHub repository if you find that it has expired.

Andreas Robinson

email: [email protected]

Felix Järemo Lawin

email: [email protected]