A camera watches a breaker, recognises the move, and a brand avatar dances along, with a live score.

breakdancing-estimation.mp4

the-problem.txt

The problem

Off-the-shelf pose models lose the body exactly when breaking gets interesting: upside down, in power moves and freezes. And a street installation can't ask for sensors on the body or depend on a network. The first test measured where the generic models fail, and that result chose the architecture.

What the house made

A full loop, camera to screen. RTMPose reads 17 joints per frame, a classifier over one-second windows says toprock, footwork or power move, and the scorer counts transitions, freezes, rotations and freezes on the beat. On screen, the brand's 3D b-girl performs the detected move, with score, combo and a charge bar. Solo and Versus modes, a leaderboard, and menus driven by raising a hand.

How it was built

Four models compared on 611 hand-labelled BRACE frames: RTMPose-m beat MediaPipe in every category and runs faster. The classifier is gradient boosting over pose, speed, inversion and torso rotation, trained on BRACE. The moves are humanoid mocap rebuilt from muscle space and retargeted to any Mixamo rig, with joint error under 0.01 mm; the brand characters came from Nano Banana → Tripo → Mixamo, with a foot-contact pass so the big-headed avatar doesn't sink into the floor. Real-space test with dancers [to measure].

measured.txt

Measured

65%PCK@0.1 · RTMPose-m

44%MediaPipe full, same frames

11.7 msper frame · M4 CPU

92%correct per segment

71%of live time = ground truth

1–2 sclassifier latency

BRACE test split (ECCV 2022), keypoints and timeline of the dataset. Footwork and power move are still confused with each other — the biggest error, and the next step: a temporal model and RTMPose-x on GPU. Musicality from pose did not separate real beats from controls (p = 0.62), so it stays off the score until we record our own sessions. Accuracy with an audience, in the physical space [to measure].

screens.txt

Screens

Waiting screen: the b-girl dances an idle on the ring under the words ENTRA NA RODA — waiting screen · 640×360 · 16 colours

live detection: flare · skeleton on the right

Player select: four Cute Crew characters as cards — player select · four brand characters

credits.txt

Credits

Concept and development: Nilo River. Visual identity: Cute Crew (2023 brand manual, 2026 tokens). Validation dataset: BRACE (ECCV 2022). Video read by the backend in the film: Bboy SNACK's power-move collection (YouTube), shown only as keypoints. Other credits [to measure]