This repository contains the code that was used to conduct experiments for a master's thesis. The goal was to detect recaps, opening credits, closing credits and previews from video files in an unsupervised manner. This can be used to automate the labeling for the skip functionality of a VOD streaming service.
The experiments done in the master's thesis were done in jupyter notebooks, but as the code in these got quite messy. I packed the used code in a python package so that it can be re-used more easily.
To try this out with Docker on your own data, copy the video files of one season to a folder in the current directory named videos
, then create a script detect.py
with the contents:
import recurring_content_detector as rcd
results = rcd.detect("videos")
print(results)
Now you can run the detector by running the docker command:
docker run -it -v $(pwd):/opt/recurring-content-detector nielstenboom/recurring-content-detector:latest python detect.py
It'll first downsize the videos using ffmpeg, then it will convert the videos to feature vectors and then the algorithm for detecting the recurring content is applied, the results
variable contains the intervals of all the recurring parts.
To install the package, do the following steps (assuming you have an anaconda setup with python 3.6) if you want to run the program with the default settings:
git clone https://github.com/nielstenboom/recurring-content-detector.git
cd recurring-content-detector
conda install faiss-cpu -c pytorch
pip install mkl
# optional step: change parameters in recurring_content_detector/config.py
pip install .
It is also possible to build a docker container that does all the steps for you with the Dockerfile in the directory.
Make sure ffmpeg is in the PATH variable and that tensorflow (GPU version preferably) is installed.
Run pip uninstall recurring-content-detector
to uninstall the package.
With this code it is possible to detect recaps, opening credits, closing credits and previews in video files from a TV-show unsupervised up to a certain extent. The following image gives a schematic overview of how it works:
You can run the detector in a python program in the following way:
import recurring_content_detector as rcd
rcd.detect("/directory/with/season/videofiles")
This will run the detection by building the color histogram feature vectors. Make sure the video files you used can be sorted in the right alphabetical order similar as to when they play in the season! So episode_1 -> episode_2 -> episode_3 -> etc.. You'll get weird results otherwise.
The feature vector function can also be changed:
# options for the function are ["CNN", "CH", "CTM"]
rcd.detect("/directory/with/season/videofiles", feature_vector_function="CNN")
This will CNN vectors, which are a bit more accurate but take much longer to build.
The detect
function has many more parameters that can be tweaked, the defaults it has, are the parameters I got the best results with on my experiments.
def detect(video_dir, feature_vector_function="CH", annotations=None, artifacts_dir=None,
framejump=3, percentile=10, resize_width=320, video_start_threshold_percentile=20,
video_end_threshold_seconds=15, min_detection_size_seconds=15):
"""
The main function to call to detect recurring content. Resizes videos, converts to feature vectors
and returns the locations of recurring content within the videos.
arguments
---------
video_dir : str
Variable that should have the folder location of one season of video files.
annotations : str
Location of the annotations.csv file, if annotations is given then it will evaluate
the detections with the annotations.
feature_vector_function : str
Which type of feature vectors to use, options: ["CH", "CTM", "CNN"], default is color histograms (CH)
because of balance between speed and accuracy. This default is defined in init.py.
artifacts_dir : str
Directory location where the artifacts should be saved. Default location is the location defined
with the video_dir parameter.
framejump : int
The frame interval to use when sampling frames for the detection, a higher number means that
less frames will be taken into consideration and will improve the processing time.
But will probably cost accuracy.
percentile : int
Which percentile of the best matches will be taken into consideration as recurring content.
A high percentile will means a higher recall, lower precision.
A low percentile means a lower recall and higher precision.
resize_width: int
Width to which the videos will be resized. A lower number means higher processing speed but
less accuracy and vice versa.
video_start_threshold_percentile: int
Percentage of the start of the video in which the detections will be marked as detections.
As recaps and opening credits only occur at the first parts of video files, this parameter can alter
that threshold. So putting 20 in here means that if we find recurring content in the first 20% of
frames of the video, it will be marked as a detection. If it's detected later than 20%, then the
detection will be ignored.
video_end_threshold_seconds: int
Number of seconds threshold in which the final detection at the end of the video should end for it
to count. Putting 15 here means that a detection at the end of a video will only be marked as a
detection if the detection ends in the last 15 seconds of the video.
min_detection_size_seconds: int
Minimal amount of seconds a detection should be before counting it as a detection. As credits &
recaps & previews generally never consist of a few seconds, it's wise to pick at least a number
higher than 10.
returns
-------
dictionary
dictionary with timestamp detections in seconds list for every video file name
{"episode1.mp4" : [(start1, end1), (start2, end2)],
"episode2.mp4" : [(start1, end1), (start2, end2)],
...
"episode10.mp4" : [(start1, end1), (start2, end2)]
}
"""
If you want to quantitively test out how well this works on your own data, fill in the annotations file and supply it as the second parameter.
rcd.detect("directory/with/season/videofiles", annotations = "path/to/annotations.csv")
When the program is done, it will give an example with the same outline as the example below:
Detections for: episode1.mp4
0:01:21.600000 - 0:02:20.880000
0:02:49.320000 - 0:03:15.480000
Detections for: episode2.mp4
0:00:00 - 0:01:16.920000
0:01:38.040000 - 0:02:37.440000
Detections for: episode3.mp4
0:00:00 - 0:01:27.840000
0:01:51.120000 - 0:02:50.760000
0:42:26.400000 - 0:42:54
Total precision = 0.862
Total recall = 0.853
There's a few tests in the test directory. They can also be run in the docker container, make sure you created a videos
directory with some episodes in it:
docker run -it -v $(pwd):/opt/recurring-content-detector nielstenboom/recurring-content-detector:latest python -m pytest -s
- https://github.com/noagarcia/keras_rmac for the CNN vectors
- https://github.com/facebookresearch/faiss for the efficient matching of the feature vectors
If you use and like my project or want to discuss something related, I would β€οΈ to hear about it! You can send me an email at [email protected].