Credits

Hand-sign recognition uses a model trained on hand-landmark annotations from HaGRIDv2 (Nuzhdin, Nagaev, Sautin, Kapitanov & Kvanchiani, 2024; github.com/hukenovs/hagrid), used under a variant of CC BY-SA 4.0. Changes: landmarks only, classes remapped and filtered. Our model is released under CC BY-SA 4.0.

The trained model files (throw_mlp.onnx and throw_mlp.weights.bin) are released under CC BY-SA 4.0 with the attribution above.

Hand landmarks come from MediaPipe Hand Landmarker by Google (the hand_landmarker.task model and the @mediapipe/tasks-vision library), used under the Apache License 2.0. Source: github.com/google-ai-edge/mediapipe.

Big Shoulders Stencil by Patric King (Copyright 2019 The Big Shoulders Project Authors) and Libre Franklin by Impallari Type (Copyright 2020 The Libre Franklin Project Authors), both used under the SIL Open Font License 1.1.

Background art and illustrations made with ChatGPT image generation, directed by the owner.

“Flowerbed Fields [Loop]” by Zane Little Music (OpenGameArt.org), CC0.

Games with a friend use PeerJS for the connection between the two devices, set up through the public PeerJS server; supabase-js and Supabase Realtime for the rooms; and zod to check every game message. All three libraries are used under the MIT License. The public STUN servers of Google (stun.l.google.com, stun1.l.google.com) and Cloudflare (stun.cloudflare.com) help the two devices find a direct route to each other. Cloudflare Realtime TURN is a relay that passes the video and voice on, encrypted, when the two devices can't reach each other directly.

Howzat never records or stores your video or voice. Against the computer your video never leaves your device; in a game with a friend your video, and your voice if you turn it on, go to your friend's device, directly or through that relay.

Page visits are counted anonymously. Your video is never part of it.