Feature Extraction Using I3D Model

This repository is created by a complete reference from https://github.com/piergiaj/pytorch-i3d, for a custom video dataset feature extraction purposes.

Note the Following Essentials Steps

This repo expects the frames to be dumped at the desired FPS beforehand.

The following efficient FFMPEG script may be used

# Check if ffmpeg is installed
if ! command -v ffmpeg &> /dev/null; then
    echo "ffmpeg is not installed. Please install it before running this script."
    exit 1
fi

# Input directory containing .mkv videos
input_dir="path_to_input_videos"

# Output directory for extracted frames
output_dir="path_where_frames_are_to_be_stored"

# Iterate through each .mkv video
for video_path in "$input_dir"/*.mkv; do
    # Get the video filename without extension
    video_filename=$(basename "$video_path")
    video_name="${video_filename%.mkv}"

    # Create a directory with the video's name in the output directory
    video_output_dir="$output_dir/$video_name"
    mkdir -p "$video_output_dir"

    # Extract frames from the video at 1 FPS and save them in the output directory
    ffmpeg -i "$video_path" -vf "fps=1" "$video_output_dir/frame_%04d.jpg"
done

echo "Frame extraction complete."

Once frames extracted, run python create_json.py which will create a JSON mapping expected by the code as file vid_info.json.
The models may be downloaded and placed in models folder, from the original repo.
python extract_features.py -mode rgb -load_model models/rgb_imagenet.pt -root /ssd_scratch/cvit/varun/output_frames_cmd/ -gpu 0 -save_dir /ssd_scratch/cvit/varun/i3d_feats

varungupta31/i3d_feature_extraction

Feature Extraction Using I3D Model