Using Jev to decide caption placement: one question per line
The author feeds MediaPipe face boxes, mouth keypoints and lip-audio sync into Jev, turning them into two-choice decisions per caption line: which face or offscreen voice owns it, and where to break a long line, following Netflix caption guidelines.
Jev experiment: Caption positioning Captioning is solved. Deciding where each caption goes is still manual work. MediaPipe gives 1. face boxes, 2. a mouth keypoint, and 3. a Face Mesh check that rejects the back of a head. The deterministic code turns that into one number: how well each mouth moves in time with the audio. Jev turns it into a decision. Two-choice questions per line: 1. which face speaks it — options are each face, or offscreen 2. where to break a line that runs long This follows the best practices of Netflix's Timed Text Style Guide