Training, Dataset creation and Benchmarking repository for Gygli convolutional neural network (https://arxiv.org/pdf/1705.08214.pdf) shot boundary detector
The model in its pre-trained form is availale in the model directory if required and checkpoints for 30 training epochs. The model is benchmarked using the RAI benchmark and obtains recall, precision and F1 score of 0.741, 0.743 and 0.742
This repo is an extension of the repo found at https://github.com/oladeha2/shot_boudary_detector and is for those who desire to further train the model or are in need of a frame based dataset for training of their own shot boundary detectors. The model is implemented using pytorch.
It contains fully commented scripts that do the following:
- Create the training and validation set, which consists of approximately 980,000 videos based on the content of two YouTube playlists https://www.youtube.com/playlist?list=PLxf1dxhJ3H9orru0qzPy1j5VDa41c4x7Z and https://www.youtube.com/playlist?list=PLxf1dxhJ3H9pzLItmYdDeBQa0RE8zmsC3. Videos should be added to the second playlist for attempted expansion (create_data_set.py)
- Training the network using the above dataset (train.py)
- Testing the trained model. This outputs the precision, recall and F1 scores for each video and overall average (test_video_rai.py). Frame content for all ten videos of the test set are available in this repo for testing.
The remaining files are custom made utilities.
The following technologies are required for successful usage of all code in the repo
- Pytube
- Moviepy
- Numpy
- Pandas
- Pytorch
- Cuda
- Matplotlib
- Pillow
- BeautifulSoup
Contact ameenbello@gmail.com for any questions regarding the contents of the repo