100,000 labeled video frames across 4 hand-object contact classes and 2 hand-side categories. The data includes bounding box coordinates for hands and interacted objects alongside labels for self-contact, another person contact, and portable object contact.
Use Cases
- Train a hand-object interaction detector using the hand and object bounding box coordinates
- Classify the physical relationship between hands and items using the contact_state labels
- Distinguish between left and right hands in egocentric or third-person views using the hand_side attribute
Strengths
- 100,000 frames with bounding box annotations for hands and objects
- Hand-side labels (left/right) for every detected hand instance
- Four-way classification of contact states: No Contact, Self Contact, Another Person, and Portable Object
- Sourced from 100 days of YouTube video footage covering daily human activities