IMPACT-Scribe: Interactive Temporal Action Segmentation with Boundary Scribbles
Researchers have created a novel system called IMPACT-Scribe aimed at improving the labeling of procedural activity videos over time. This framework integrates several advanced techniques, including uncertainty-aware boundary scribble supervision and cost-aware query planning, to foster better teamwork between people and machines. Notably, every correction made enhances future interactions rather than standing alone. Findings from experiments and a human study indicate that the quality of labeling, boundary precision, and the dynamics of human-machine collaboration all progress with use. Additionally, the code for this innovative framework is set to be made available to the public.
Key facts
- IMPACT-Scribe is a correction-driven framework for dense temporal labeling.
- It uses uncertainty-aware boundary scribble supervision.
- It incorporates local proposal modeling and cost-aware query planning.
- The system features structured propagation and correction-driven adaptation.
- Each correction improves future human-machine collaboration.
- Experiments and a human study show improved labeling quality per effort.
- Boundary accuracy is enhanced with the framework.
- The code will be publicly available.
Entities
—