Extract Video Feature

(I miss spell "extract" and this repository name is not correct spelling) extruct video feature from C3D.

The output follows the MAC.

Requirements

• Python 3.7
• PyTorch 1.10
• torchfile 0.1.0
• others

Model

• C3D

model is refered to DavideA/c3d-pytorch.

and you can download pretrained weight(Sports-1M) of C3D from c3d.pickle and pretrained weight(Kinetics) of C3D from c3d-sports1m-kinetics.t7.

Video Format

• mp4

Option

• --root_dir (str) : give a directory path that have videos you want to extract feature.
• --frame_unit (int) : specify frame length input to model at once. Default: 16.
• --overlap (float) : specify frame overlap percentage. If you specify 16 to frame_unit and 0.5 to overlap, overlap frame is 8 frame(16*0.5), so start frame are 1,9, 17, 25,... Default : 0.0.
• --out_dir (str) : specify the path to put output feature. Default: ./output
• --gpu_id (int) : specify GPU ID that you use. Default: 0.
• --pretrained_path (str) : specify pretrained weight path.
• --pretrain (str) : specify pretrain dataset. you can choose sport-1m or kinetics. Default: kinetics.
• --verbose (flag) : If you add this option, print saved file names.

Usage

1. download weight.

python3 main.py --root_dir <video path> --frame_unit <unit> --overlap <Overlap> --gpu_id <GPU ID> --pretrained_path <path of c3d.pickle>

example

python3 main.py --root_dir /mnt/dataset/Charades --frame_unit 16 --overlap 0.5 --pretrained_path ./c3d.pickle --out_dir ./feature --verbose

Save format

.npy