====================================

MOVES is a self-supervised means of learning visual features useful for segmentation from arbitrary collections of video. At inference time, MOVES requires only an image to produce visual features and a segmentation mask.

Additionally, we demonstrate how one can use pseudolabel segmentation masks as a means of discriminating between people and the objects they use. One could also use robot segmentation masks to similarly apply this addition to robotic manipulation video.

To learn more, visit the website: https://relh.github.io/moves/

Your mileage will vary, but something like this on a GPU machine should work. Use conda if you haven't switched to mamba:

mamba install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
mamba install -c rapidsai -c conda-forge -c nvidia \
              rapids=23.10 cuml=23.10 python=3.10 cuda-version=11.8
pip install -U openmim
mim install mmcv
pip install -r requirements.txt
mim download mmflow --config raft_8x2_100k_flyingthings3d_sintel_368x768

Note in the above you will have to separetly install PyTorch, RapidsAI, and MMFLOW. The last line downloads an optical flow model.

As an example, lets try an arbitrary dataset, how about UCF101.

1. Visit the webpage and download the UCF101.rar
2. Run unrar e UCF101.rar.
3. Put the videos in a folder, like UCF101/
4. Run CUDA_VISIBLE_DEVICES=0 python pseudolabeller.py --video_path=./UCF101/, this will:

• extract frames
• detect people in the frames using ternaus
• cache forward and backwards optical flow between subsequent frames

5. Run CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python main.py --name ucf101demo --model hrnet --target ddaa --people --port 23232 --train --inference
6. Inside the experiments folder will be your experiment, it will have a index.html and test_index.html file with saved outputs. We use apache so we make a symlink to public html to view results.

Caution, the pseudolabeller will use all the GPUs it sees.

mkdir UCF101/
cd UCF101/
wget https://www.crcv.ucf.edu/data/UCF101/UCF101.rar
unrar e UCF101.rar
cd ..
CUDA_VISIBLE_DEVICES=0 python pseudolabeller.py --video_path=./UCF101/
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python main.py --name ucf101demo --model hrnet \
                     --target ddaa --people --port 23232 --train --inference

If everything has worked, you should see results pages that look like this:

And then this for inference, where the HDBSCAN segments (labelled as clusters) are the output.

If you find our method or this repo helpful, please consider citing our conference paper:

@inproceedings{higgins2023moves,
  title={MOVES: Manipulated Objects in Video Enable Segmentation},
  author={Higgins, Richard EL and Fouhey, David F},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6334--6343},
  year={2023}
}