Gaze direction determination of opponents and teammates is a very important ability for any soccer player, human or robot. However, this ability is still not developed in any of the RoboCup soccer leagues. We aim at reverting this situation by proposing a gaze direction determination system for robot soccer; the system is designed primarily for the four-legged league, but it could be extended to other leagues. This system is based on a robot-head pose detection system, consisting on two main processing stages: (i) computation of scale-invariant local descriptors of the observed scene, and (ii) matching of these descriptors against descriptors of robot-head prototypes already stored in a model database. After the robot-head pose is detected, the robot gaze direction is determined using a head model of the observed robot, and the current 3D position of the observing robot camera. Experimental results of the proposed approach are presented.