Benchmark
Shotwell-Bench
Can a model tell when one action ends and another begins in first-person video, then describe what each hand did?
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Shotwell selected episodes from WGO-Bench with clearly visible actions and annotated where multiple ways to split an action are valid.
Methodology
The segmentation task accepts a range of times for each boundary and allows optional splits. A second task supplies the segments and uses a language-model judge to check descriptions for the required actions, objects and hands. Each task has its own F1 score. The selected clips favor actions that annotators can distinguish clearly.