Gemini 4 Carbon Video Test: How Close Is Its Timing?

A test of Gemini 4 Carbon's ability to time an archery arrow release and draw start in video, compared against manual frame-by-frame annotation.

Key Takeaways

Gemini 4 Carbon was reasonably close on sharp, instantaneous events like arrow release (about two frames off).
Gradual, blended events like the start of a draw are significantly harder to time accurately for both AI and humans.
Human annotators can be inconsistent on fuzzy moments, so explicit labeling rules are essential for video datasets.
Defining exact start and end criteria for each phase before labeling improves consistency and reliability.

What the video covers

The video tests whether Google's Gemini 4 Carbon model can accurately identify event timing in a 14-second archery clip.
The model estimated the arrow release at around 11 seconds; manual annotation placed it at 10.919 seconds, about 81 milliseconds (roughly two frames) apart.
For the start of the draw, the model said 6 seconds while manual marking gave 5.29 seconds, a much larger discrepancy.
The presenter explains that releases happen instantaneously and are easier to pinpoint, whereas draws blend gradually into preceding motion, leaving no single obvious frame.
A second clip was marked twice by the presenter to test human consistency: release marks differed by just 7 milliseconds, while draw-start marks differed by 83 milliseconds.
The results suggest that ambiguous, gradual transitions are the hardest for both the AI model and human annotators.
The video frames the outcome as a lesson for video labeling: define precisely where each phase begins before annotating.
The content is an informal experiment rather than a rigorous benchmark, based on a small number of clips.
The channel, OpenCV University (Learn OpenCV), presents it as part of a series of computer vision experiments.

How close was Gemini 4 Carbon's estimate of the arrow release timing?

Gemini 4 Carbon estimated the arrow release at around 11 seconds, while the presenter's manual mark was 10.919 seconds. The difference of about 81 milliseconds is roughly two frames, which the presenter considers close.

Why was the draw start harder to time than the release?

The presenter explains that a release happens in one instant, while a draw blends into the movement before it, so there is no single obvious frame. This made the draw start far more ambiguous, with the model estimating 6 seconds versus the manual 5.29 seconds.

What advice does the video give for video labeling projects?

The presenter recommends deciding exactly where each phase begins before starting annotation. The video shows that fuzzy, gradual moments are difficult for both models and humans, so clear definitions help keep labels consistent.