Authors:
Kevin Zhu
;
Alexander Wong
and
John McPhee
Affiliation:
University of Waterloo, Canada
Keyword(s):
Chess, Automated Digitization, Computer Vision, Deep Learning.
Abstract:
A digitized chess match offers chess players a convenient way to study previous matches. However, manually recording a large number of matches can be laborious, while automated methods are usually hardware-based, requiring expensive chessboards. Computer vision provides a more accessible way to track matches from videos. However, current vision-based digitizers are often evaluated on images captured by cameras placed directly above a chessboard, and performance suffers when the camera angle is lower, limiting their applicability. Motivated to develop a more practical solution, we introduce VICE, a view-invariant chess estimator to digitize matches from camera angles not seen during training. Due to its small model size and computational efficiency, VICE is suitable for mobile deployment. By rearranging the framework for chess detection and incorporating prior information from chess and basic geometry, we simplify the chess estimation problem and mitigate the challenges that current c
hess digitizers struggle with, such as occlusion. We combine the board localization and chess piece detection phases of classical two-step chess estimation to develop a prototype for the first single-step chess digitizer. We show that, with minimal training data, our prototype can infer moves from camera angles that current chess digitizers cannot, while being much smaller in size.
(More)