Note: In this context, a video is segmented into multiple "clips," and a feature vector is extracted from each "clip."
-
video_dir_path
: The directory (absolute path) where the input videos are stored. All videos in this directory will be processed. -
out_path
: The directory (absolute path) where the extracted features will be stored. -
batch_size
: The number of "clips" that get fed into the model at each iteration. -
frame_window
: The number of frames in a "clip." -
stride
: The number of frames between the first frame of two neighboring "clips." -
use_remote
: Whether you want to use a remote checkpoint or a local one for the model weights. -
ckpt
: The name/path of the model checkpoint (can be either local or remote, depending onuse_remote
). -
side_size
: The size to which the shorter side of the input video will be resized. -
crop
: You can choose between "three_crops" and "center." This option determines how the input frames will be cropped. For more customization, please go tomodels/model_name.py
and modify theget_transform
method. -
crop_size
: The size of the crop mentioned above. -
model_module_str
: The location of the Python file where you implement the model (default: undermodels/
).