Data Curation

This repository provides simple tools for filtering and extracting dynamic video sequences using optical flow magnitude.

Originally developed for the VoxCeleb dataset, but applicable to any frame-level video dataset with consistent naming (e.g., video_name/frame.png).

Scripts Overview

• clip_level_filtering.py
  Extracts dynamic clips using a sliding window and merges overlapping segments.
  Output: filtered_dynamic_clips.csv

• video_level_filtering.py
  Filters videos with low motion based on average optical flow score.
  Output: filtered_videos.csv, optical_flow_scores.csv

• visualize_threshold_distribution.py
  Plots the number of videos above each threshold to guide cutoff selection.
  Output: optical_flow_threshold_plot.png

Input Format

<root_dir>/
├── id0001/
│   ├── 000.png
│   ├── 001.png
│   └── ...

Run scripts

Edit the ROOT_DIR and THRESHOLD values in each script, then run:

python clip_level_filtering.py
python video_level_filtering.py
python visualize_threshold_distribution.py