IMPACT-HOI: Mixed-Initiative Framework for Egocentric Video Annotation
IMPACT-HOI is a teamwork approach for annotating videos that focus on how people interact with objects. It views the annotation process as a way to gradually clarify events that start with some initial information. A smart controller helps by offering direct questions, suggestions verified by people, and careful completions, while also considering what the annotator does and the evidence's quality. To prevent automated changes from overriding human decisions, a careful execution system with rollback capabilities is used. In a study with 9 participants, there was a 13.5% reduction in manual tasks, a 46.67% match rate for events, and no violations in confirmed fields. The code will be made publicly available.
Key facts
- IMPACT-HOI is a mixed-initiative framework for egocentric procedural video annotation.
- It constructs structured event graphs for Human-Object Interactions (HOI).
- The framework uses a trust-calibrated controller for query selection.
- A risk-bounded execution protocol with atomic rollback preserves human decisions.
- User study with 9 participants showed 13.5% reduction in manual actions.
- Event match rate was 46.67%.
- Zero confirmed-field violations occurred under the studied protocol.
- Code will be made publicly available.
Entities
—