A coach watching an eight sees one seat at a time. By the time they have watched all eight, the piece is over. The feedback that actually changes a rower is specific, immediate and repeatable, and a human standing on a launch can deliver at most one of those three.
The equipment that solves this exists. It lives in sports-science labs, costs more than most clubs' annual budget, and requires the rower to come to it. Meanwhile every coach on the water is already carrying a camera good enough to do the job, if something on the other end knew what to look for.
The measurement was never the hard part. Getting it to a rower on a Tuesday morning was.
Pose keypoints come out of the video frame by frame. From those the pipeline segments the piece into individual strokes, then splits each stroke into catch, drive, finish and recovery. Every stroke is then run against nine checks, grouped into the three things a coach actually says out loud: the power chain, the body positions, and the rhythm.
Each check returns a proportion, not a verdict. Legs then back then arms firing in the right order on 31% of strokes is a different conversation from 95%, and the difference is what a rower can act on. A camera that is not square to the boat has its angles corrected before any of this runs, because a launch panning across a crew is the normal case, not the exception.
A launch is moving, panning, and never perpendicular to the boat. Every body angle measured off that footage is wrong in the same direction. Swing the camera below and watch what a naive pipeline would report.
Swing the camera off square. The dashed figure is the rower's true shape; the solid one is what the lens records. Every angle on the right is re-measured from those pixels.
The sport this came out of. World #6, and three years of 5:30 alarms.
The other place I cared more about the honesty of a number than its size.
Alliance selection for FTC, built on the same instinct: measure it, then decide.