Your pull-up form, judged by a pose model with no mercy

What it does Each demo takes a video of a real activity and produces numbers you’d normally need a coach for: rep counts and per-rep timing for chin-ups, pull-ups, push-ups, and deadlifts; running cadence and knee shape at foot strike; a dance-sync similarity score; basketball make/miss calls with a shot chart; and rock-climbing route analysis that tracks which holds you used and in what order. Most of it runs on vitpose-plus-large for pose, with sam3.1 segmentation where objects (ball, rim, holds, plates) matter.

The interesting bit The pull-up demo doesn’t just count — it grades each rep as full or partial against your own calibrated dead hang, which is a smarter reference than any fixed threshold. And rock_climbing_3d reconstructs a boulder problem in actual meters from an iPhone LiDAR capture, wall, holds, and climber included.

Key highlights

• Nine demos spanning gym lifts, running, climbing, basketball, and dance
• Per-rep quality grading, not just counting: range of motion, straight back, rounded-back probability
• Climbing route comparison across attempts, plus a LiDAR-based 3D reconstruction with metric measurements
• Most demos work in a browser app with no API key (anonymous, rate-limited tier)
• Apache-2.0 licensed, with per-project READMEs

Caveats

• rock_climbing_3d, push_ups, and basketball aren’t available in the browser app — the 3D one needs a LiDAR capture.
• ffmpeg is required, and higher usage limits need a VLM Run API key.

Verdict Great if you want working reference implementations of pose estimation applied to messy real-world sports footage — or a cheap form coach. Skip it if you wanted a single polished product rather than nine separate experiments.