Every rep from four side-view videos, scored on depth, body line, lockout, head and knees — with the angles drawn on the video so you can check each score yourself.
Loading…
Use a pose model for scoring, and a vision-language model only for commentary. The four pose models agree on every rep in three of four sessions (60 / 60 / 60). In the backlit fourth they agree on 50, and only YOLO double-counts. RTMPose-x is the best single model: it agrees most with the others, shakes least among the accurate ones, and runs at ~60 ms/frame. The ensemble (median of all four) is the safest default.
Gemini 4 Argon finds 90–100% of the reps and writes genuinely useful coaching — it spotted the piked top and the "worming" ascent. But its numbers are flat and miscalibrated, and it can't draw proof. Gemini 3.8 Flash missed 35 seconds of real push-ups.
No hand-labelled keypoints exist, so "accuracy" here means agreement with the other models, frame-to-frame steadiness, rep counts against an adjudicated set (reps ≥ 3 of 4 pose models agree on, plus eyeballed rulings on every dispute), and the visual audit below. Click a row to view that approach.
Hard frames (transitions, the bottom of a rep, backlight). Blue = left side, red = right side, green dot = nose.