Optical Flow Extraction for UCF101

This is the torch code, extracting optical flow of videos in UCF 101 dataset. This repo includes and uses torch-ffmpeg by Carl Vondrick for FFmpeg usage. Optical extraction part is largely based on the code: VisionLabs/torch-opencv/demo/cuda/optical_flow.lua. Optical flow will be saved as jpeg file, using first 2 channels to store the optical flow values.

Requirements

1. install cutorch and cv
2. download UCF101

Usage

Extract image frames from UCF101 videos.

th video2images.lua -dataset <path-to-UCF-101> -save <save directory>

Generate optical flows from the extracted video frames.

th img2optflow.lua -dataset <frame directory> -save <optical flow directory>

Sample Code.

th video2images.lua -dataset UCF/UCF-101 -save UCF/frames
th img2optflow.lua -dataset UCF/frames -save UCF/optflow