Point-of-No-Return (PNR) localization using SlowFast with Perceiver backbone.
Use the following command to install the required packages:
pip install -r requirements.txt
Change the following depending on your dataset configuration:
Open the file default.yaml
Then, modify the following entries with the appropriate paths:
Under the DATA category
Change
VIDEO_DIR_PATH: to the path to the Ego4D videos
ANN_PATH: to the path to your annotation file
SPLIT_PATH: to the path to the folder with split files
CLIPS_SAVE_PATH: to the path to the folder that contains frames extracted from Ego4D videos or the path where you want to
store the extracted frames
VIDEO_LOCATIONS_CSV: to the path to a CSV file with paths to videos
NO_SC_PATH: to the path to the folder that contains frames extracted from Ego4D for videos without state change
NO_SC_SPLIT_PATH: to the path to JSON file containing the splits for videos without state change
Under MISC
Change
NUM_GPUS: to the GPUS that will be used for training
Then, run the following command
To train the model run the following command from the project root
python run.py --cfg configs/default.yaml
To test the model run the following command from the project root
python run.py --cfg configs/default_test.yaml