videosum 0.0.8

Python module to summarise a video into a collage.

Description

Code for the videosum Python package. Given a video file, this package produces a single-image storyboard that summarises the video.

Citation

If you use this code in your research, please cite the paper:

@article{GarciaPerazaHerrera2023,
	author = {Luis C. Garcia-Peraza-Herrera and Sebastien Ourselin and Tom Vercauteren},
	title = {VideoSum: A Python Library for Surgical Video Summarization},
        journal = {arXiv preprint arXiv:2303.10173},
	year = {2023}
}

Supported platforms

• Ubuntu >= 20.04
• Python >= 3.10

Use the package from a Docker container

If you want to quickly try the package, this is the easiest way, provided that you have Docker installed with GPU support. If you do not have it, you can follow this tutorial to install Docker with GPU support. Alternatively, you can install the videosum package from pip or from source, following the next sections of this README.

1. Download repository:

   $ git clone https://github.com/luiscarlosgph/videosum.git
   $ cd videosum/docker
   

2. Build Docker image:

   $ docker build --build-arg USER=$(whoami) --build-arg UID=$(id -u) --build-arg GID=$(id -g) -t luiscarlosgph/videosum:latest .
   

3. Run videosum Docker container:

   $ docker run --volume $HOME:/mnt/user_home --name videosum --runtime nvidia luiscarlosgph/videosum:latest &
   

   The --volume argument mounts your home directory (outside the container) into the directory /mnt/user_home inside the container. This is so that you can access the videos that need to be summarised from within the container, and also save the output storyboards in a folder that is accessible from outside the Docker container.

4. Get a terminal on the container:

   $ docker exec --user $(whoami) --workdir $HOME -it videosum /bin/zsh
   

5. Summarise the test video:

   $ python -m videosum.run --input /opt/videosum/test/data/video.mp4 --output /mnt/user_home/storyboard.jpg --fps 2 --nframes 16 --height 1080 --width 1920 --algo inception
   

   After running this command you should have the summary of the test video saved as storyboard.jpg in you home directory.

Install dependencies

If you are not using the videosum Python package via Docker (as explained in the previous section), but you are planning to install it natively in your system, you must install all the dependencies detailed in this section.

• Python version 3.10: works well with faiss version 1.7.4. You can find a howto guide to install this version of Python here.

• PyTorch version 1.12.1+cu116, you can install it running:

  $ pip install torch==1.12.1+cu116 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116`
  

• ffmpeg: in Ubuntu/Debian, run $ sudo apt install ffmpeg to install it.

• swig: in Ubuntu/Debian, run $ sudo apt install swig to install it.

• faiss-gpu version 1.7.4: you can find a howto guide to install it here.

A good way to see all the commands that need to be executed to install all the dependencies is by checking the Dockerfile.

Install with pip

$ python3 -m pip install videosum --user

Install from source

$ git clone https://github.com/luiscarlosgph/videosum.git
$ cd videosum
$ python3 setup.py install --user

Run video summarisation on a single video

$ python3 -m videosum.run --input video.mp4 --output collage.jpg --nframes 100 --height 1080 --width 1920 --algo time

Options:

• --input: path to the input video file.
• --output: path where the output collage will be saved.
• --nframes: number of frames that you want to see in the collage image.
• --height: height of the collage image.
• --width: width of the collage image.
• --time-segmentation: set it to either 0 or 1. If 1, the clustering results are displayed in a bar underneath the collage (i.e. the columns of the bar represent the frames of the video, and the colours represent the clustering label).
• --fps: number of frames you want to read per second of video, used to downsample the input video and have less frames to describe and cluster.
• --time-smoothing: weight in the range [0.0, 1.0] that regulates the importance of time for clustering frames. A higher weight will result in a segmentation of frames over time closer to that of the time method.
• --processes: number of processes to use when summarising a folder of videos.
• --metric: set it to True if you want to compute the Frechet Inception Distance between the frames in the summary and the frames in the original video.
• --algo: algorithm used to select the key frames of the video.
  • time: evenly spaced frames are selected.
  • inception: k-medoids clustering (l2-norm metric) on InceptionV3 latent space vectors.
  • uid: in order to compute the distance between two images their InceptionV3 latent space vectors are computed, a univariate Gaussian is estimated for each of the two latent vectors, and the 2-Wasserstein distance is computed between the two Gaussians and used as clustering metric for k-medoids. uid stands for Univariate Inception Distance.
  • scda: k-medoids clustering (l2-norm metric) on SCDA image descriptors (Wei et al. 2017 Selective Convolutional Descriptor Aggregation for Fine-Grained Image Retrieval). In this package we use InceptionV3 as opposed to VGG-16, which was the model used by Wei et al.

Run video summarisation on multiple videos

Pointing the command line parameter -i or --input to a folder of videos is enough. In this case, the path indicated by -o or --output will be used as output folder, each video summary will have the same filename as the video but a .jpg file extension.

The parameter --processes allows you to select the number of videos to summarise in parallel. This is necessary because some of the summarisation methods use GPU memory, which is typically a limiting factor. If the number of processes is too high you might get a CUDA out of memory error.

$ python3 -m videosum.run -i <input_folder> -o <output_folder> -n 16 --width 1920 --height 1080 -a inception --fps 1 --time-segmentation 1 --processes 5

Exemplary code snippet

import cv2
import videosum

# Choose the number of frames you want in the summary
nframes = 100

# Choose the dimensions of the collage
widtth = 1920
height = 1080

# Choose the algotrithm that selects the key frames
algo = 'inception'  # The options are: 'time', 'inception', 'uid', 'scda'

# Create video summariser object
vs = videosum.VideoSummariser(algo, nframes, width, height)

# Create collage image
im = vs.summarise('video.mp4')

# Save collage to file
cv2.imwrite('collage.jpg', im)

# Retrieve a list of Numpy/OpenCV BGR images corresponding to the key frames of the video
key_frames = vs.get_key_frames('video.mp4')       

# Print the video frame indices of the key frames, available after calling summarise() or get_key_frames()
print(vs.indices_)

# Print the video frame cluster labels, available after calling summarise() or get_key_frames()
print(vs.labels_)

Exemplary result

The storyboards have a bar underneath that is produced when the --time-segmentation 1 option is passed. This bar shows how frames have been clustered over time, with a colour for each cluster, and black vertical lines representing the key frames.

• Exemplary video: here

• Summary based on time algorithm:

$ python3 -m videosum.run --input test/data/video.mp4 --output test/data/time.jpg --nframes 16 --height 1080 --width 1920 --algo time --time-segmentation 1

• Summary based on inception algorithm:

$ python3 -m videosum.run --input test/data/video.mp4 --output test/data/inception.jpg --nframes 16 --height 1080 --width 1920 --algo inception --time-segmentation 1

• Summary based on uid algorithm:

$ python3 -m videosum.run --input test/data/video.mp4 --output test/data/uid.jpg --nframes 16 --height 1080 --width 1920 --algo uid --time-segmentation 1

• Summary based on scda algorithm:

$ python3 -m videosum.run --input test/data/video.mp4 --output test/data/scda.jpg --nframes 16 --height 1080 --width 1920 --algo scda --time-segmentation 1

Run unit testing

$ python3 setup.py test

Run timing script

$ python3 -m videosum.timing 

Method	Time for a 1h video sampled at 1fps
time	13s
inception	86s
uid	216s
scda	74s

Use this package as an unsupervised spatial feature extractor

If you have 2D RGB images and you want to obtain a feature vector for them, you can do so like this:

1. Install videosum Python package:

   $ pip install videosum --user
   

2. Extract feature vectors for your images:

   import cv2
   import videosum
   
   # Read a BGR image from file
   im = cv2.imread('test/data/test_time.png', cv2.IMREAD_UNCHANGED)
   
   # Extract latent space spatial feature vector for the image
   model = videosum.InceptionFeatureExtractor('vector')
   vec = model.get_latent_feature_vector(im)  # Here you can pass an image (H, W, 3) 
                                              # or a batch of images (B, H, W, 3)
   
   # Print vector dimensions
   print(vec)
   print('Shape:', vec.shape)
   

   The output:

   [0.34318596 0.11794803 0.04767929 ... 0.09731872 0.         1.1942172 ]
   Shape: (2048,)
   

Author

Luis Carlos Garcia Peraza Herrera ([e-mail]), 2022-present.

License

This code repository is shared under an MIT license.