✊ Rock-Paper-Scissors vs. the Net
Show one hand to the camera. First to 3 rounds wins. The side panel shows the model's live softmax — watch it get confident right before it reads your throw.
First to 3 wins. Hold your throw steady for the last second of the countdown.
🧠 What just happened
The model found 21 3D keypoints on your hand, fed the geometry through a classifier, and the bars you watched were a softmax — probabilities over the model's gesture vocabulary (fist, palm, victory…), which we translate into Rock/Paper/Scissors. When the classifier hesitates on a fist seen from an odd angle, a pure-geometry fallback measures whether all four fingertips are curled into the palm in 3D — angle-proof, but only a heuristic. Fast blurry hands win rounds because the classifier needs frames of stability to converge. That latency-vs-confidence trade-off is a real production problem in every gesture product.