Supported Models

Action Recognition

• S3D (Kinetics 400)
• R(2+1)d RGB (IG-65M, Kinetics 400)
• I3D-Net RGB + Flow (Kinetics 400)

Sound Recognition

• VGGish (AudioSet)

Optical Flow

• RAFT (FlyingChairs, FlyingThings3D, Sintel, KITTI)

Frame-wise Features

• All models from TIMM e.g. ViT, ConvNeXt, EVA, Swin, DINO (ImageNet, LAION, etc)
• CLIP
• ResNet-18,34,50,101,152 (ImageNet)

Quick Start

or run with conda locally:

# clone the repo and change the working directory
git clone https://github.com/v-iashin/video_features.git
cd video_features

# install environment
conda env create -f conda_env.yml

# load the environment
conda activate video_features

# extract r(2+1)d features for the sample videos
# (defaults to printing; use on_extraction=save_numpy or save_h5 or save_pickle to save to disk)
python main.py \
    feature_type=r21d \
    device="cuda:0" \
    video_paths="[./sample/v_ZNVhz7ctTq0.mp4, ./sample/v_GGSY1Qvo990.mp4]"

# if you have many GPUs, just run this command from another terminal with another device
# device can also be "cpu"

If you are more comfortable with Docker, there is a Docker image with a pre-installed environment that supports all models. Check out the Docker support. documentation page.

Multi-GPU and Multi-Node Setups

With video_features, it is easy to parallelize feature extraction among many GPUs. It is enough to start the script in another terminal with another GPU (or even the same one) pointing to the same output folder and input video paths. The script will check if the features already exist and skip them. It will also try to load the feature file to check if it is corrupted (i.e. not openable). This approach allows you to continue feature extraction if the previous script failed for some reason.

If you have an access to a GPU cluster with shared disk space you may scale extraction with as many GPUs as you can by creating several single-GPU jobs with the same command.

Since each time the script is run the list of input files is shuffled, you don't need to worry that workers will be processing the same video. On a rare occasion when the collision happens, the script will rewrite previously extracted features.

Input

The inputs are paths to video files. Paths can be passed as a list of paths to videos or as a text file formatted with a single path per line.

Output

Output is defined by the on_extraction argument; by default it prints the features to the command line. Possible values of output are ['print', 'save_numpy', 'save_pickle', 'save_h5'].

• save_numpy / save_pickle: Saves the features in the output_path folder with the same name as the input video file but with the .npy or .pkl extension.
• save_h5: Saves features into a single HDF5 file per device in the output_path (e.g. video_features_cuda0.h5) with video path as keys (//\\ replaced by _). Structure: file.h5 / video_key (group) / feature_name (rgb, flow, fps etc). Use utils.py/inspect_h5() to explore the content.

Used in

• SpecVQGAN branch specvqgan
• BMT branch bmt
• MDVC branch mdvc

Please, let me know if you found this repo useful for your projects or papers.

Acknowledgements

• @Kamino666: added CLIP model as well as Windows and CPU support (and many other useful things).
• @ohjho: added support of 37-layer R(2+1)d favors.
• @borijang: for solving bugs with file names, I3D checkpoint loading enhancement and code style improvements.
• @bjuncek: for helping with timm models and offline discussion.
• @VivekNarula7: for adding support for .h5 output format.

Citation

If you find this project useful in your research, please consider cite:

@misc{videofeatures2020,
  title = {Video Features},
  author = {Vladimir Iashin and other contributors},
  year = {2020},
  howpublished = {\url{https://github.com/v-iashin/video_features}},
}