ego4d_selection/README.md

Ego4D Selection for Event-Driven Manipulation

This folder contains scripts to download only the Ego4D parts useful for the event-driven manipulation project.

What to Download

Priority:

1. FHO annotations: state change, PNR, hand-object interaction, active object.
2. FHO clips selected by tool/state-change keywords.

Do not download full-scale Ego4D videos first.

Prerequisite

Ego4D requires AWS credentials after license approval. Configure the default AWS profile before running the scripts.

Expected failure when credentials are missing:

RuntimeError: Could not find AWS profile 'default'.

Step 1: Download Annotations

powershell -ExecutionPolicy Bypass -File D:\event_driven\01_data\ego4d_selection\download_ego4d_annotations.ps1

Output target:

D:\ego4d_data\annotations

Step 2: Select and Download Clips

powershell -ExecutionPolicy Bypass -File D:\event_driven\01_data\ego4d_selection\select_and_download_ego4d_clips.ps1

Selection outputs:

D:\event_driven\01_data\ego4d_selection\ego4d_selected_clips.csv
D:\event_driven\01_data\ego4d_selection\ego4d_selected_video_uids.txt
D:\event_driven\01_data\ego4d_selection\selection_summary.json

Downloaded clips target:

D:\ego4d_data\clips

Why This Subset

Ego4D is mainly useful for:

• event/state-change detection
• PNR timing
• hand-object contact
• short-term active object prediction
• object-centric world model training

It is not the best primary dataset for tool YOLO detection. Keep Open Images and your own tabletop data as the main YOLO training data.