Adaptive video highlight detection

Prerequistes

• Python 3.6.9
• PyTorch 1.2.0

Dataset and setup

• Download and process the PHD-GIFs dataset.
• See config/config.py for different experiment settings and parameters.
• Fix the paths in config/config.py. We store the list of train/val/test users and their video and histories path in .json files.
• For each user, we store the video features in a .csv file and the user's history features in a .json file. For each element in the user's history, we consider the features of the segments that are indicated as highlights in the ground truth. See dataloader/make_dataloader_final_dumps.py for details and update the paths for .csv and .json files for users.
• Note that this codebase is a reimplementation. It is very likely that I may have made some mistakes during the process. However, I intend to fix them over time.
• Below I provide example training and testing commands.

Training

To train, run the following command:

python train.py --hist_net attn

Testing

To test, run the following command:

python test.py --hist -m ./checkpoints/adain-attn/checkpoint.pt --hist_net attn