Reading a physical die with a camera sounds simple: find the die and read the number. In practice, a tabletop environment introduces several computer vision problems at once.
The system first needs to identify which object is a die. It then needs to determine the visible face, cope with rotation and perspective, and decide whether the die has finished moving before accepting a result.
Different dice create different problems
A standard six sided die presents pips or numerals on flat faces. Other tabletop dice can have more faces, steeper viewing angles and much smaller markings. Materials also vary. Glossy, translucent or patterned dice may reflect light differently from opaque dice.
That means a practical dice recognition system needs to be trained and tested across the types of dice people actually use.
Timing is part of recognition
A camera may see several possible values while a die is rolling. The software therefore needs a way to determine when the roll has settled. Motion detection, frame history and confidence scoring can all contribute.
Occlusion also matters. A hand may partially cover the die, or several dice may land close together. The correct product behaviour is often to ask for confirmation rather than confidently report a low confidence result.
Why keep the physical dice?
For tabletop games, the purpose is not simply to replace arithmetic. Physical dice are part of the ritual of play. Computer vision allows the digital system to understand the result while the player still gets to roll the real object.
Doungim is being developed by Realm Kinetics around that principle. Its camera is intended to connect physical tabletop actions with the digital layer of the game.
The interesting engineering problem is therefore not "can a camera see a number?" It is whether the whole recognition pipeline can operate reliably enough to disappear into normal play.