Game Record
Introduction
Steven Scher and Ryan Crabb
ABSTRACT
We integrate in-person and on-line playing of board games such as Go. We allow a player to record an in-person game by placing their cellphone on the table next to the game board. We record the game with the cell phone's camera and automatically transcribe the game. We automatically detect the board and playing pieces, using the game's rules to accurately estimate long sequences of moves. We achieve excellent accuracy, with 9 out of every 10 games recorded without error. The game transcript is automatically uploaded in the standard SGF file format, and may be studied afterwards, shared with friends or coaches, or added to online compilations.MOTIVATION
Many board games, such as Chess and Go, are played both in-person and online. After playing a game online, the record of the game is available so that the players may discuss particularly good and bad moves with each other and with friends and teachers. This commentary often adds significantly to the social experience and skill level improvement. Reviewing a game repeatedly until every move has been understood and memorized is a common study technique. Very good players can remember a game played in person, and review them afterwards. Other players don’t have as keen memory, but they do have a cellphone.CHALLENGE
We allow a player to record their game simply by placing a cellphone camera on the table next to the board during a game. Our software on the phone automatically take pictures, finds the board and detect moves. A complete record of moves played is created, so that good and bad moves can be reexamined by those present and shared via email with others. A file in the standard SGF file format may be uploaded to the player's online account, and online game databases can be queried to find professional games facing similar decisions. TheCONTRIBUTION
Our system automatically detects the board robustly by first estimating the homography of the image to a planar grid, and searching for the transformed grid. We detect stones using an analysis-by-synthesis paradigm to avoid errors from shadows, highlights, occlusions, and minor movements of stones. We detect hands occluding the board to avoid spurious detections. Furthermore, we perform inference over our detection probabilities to leverage our knowledge of the game's rules, such as alternation of black and white stones. We achieve excellent error rates, with false negatives (failing to detect a stone present) only once in 10 games of 200 moves each, and false positives (erroneously detecting a stone) only once in 10 games.