Shot Boundary Detector based on convolutional neural network created by Gygli --> https://arxiv.org/pdf/1705.08214.pdf
To use the shot boundary detector the following python libraries are reuqired
- moviepy
- pillow (PIL)
- pytorch
- cuda
For easy use of the shot boundary detector place the video file in the same directory is possible. Clone repo and to detect the frames of the input video that are potentially part of a shot boundary run the following command while inside the repo:
- For gpu users --> python get_transition_frames_gpu.py <video_file_path>
- For cpu users --> python get_transition_frames_cpu.py <video_file_path>
The file name to svae files to must have the '.txt' extension added to it. The resulting text file located in 'predictions/<your_text_file_name>' will have the frame numbers in the input video that are part of one of the following transitions:
- Hard cut
- Crop cut
- Dissolve
- Fade In
- Fade Out
- Wipe
The model used obtains average recall, precision and f1 score of 0.743, 0.741 and 0.742 this is tested on the RAI benchmark
If the requirement is not just to use the shot boundary detector for inference but to improve the operation via traininga and test for new results the repo https://github.com/oladeha2/gygli_sbd_train_and_test can be visited
This repo contains python scripts required for local dataset creation, training and testing on the RAI videos whose frame content are already available in the repo
Contatct ameenbello@gmail.com for questions regarding repo content