High level saliency prediction for smart game balancing

Predicting visual attention can significantly improve scene design, interactivity and rendering. For example, image synthesis can be accelerated by reducing computation on non-attended scene regions; attention can also be used to improve LOD. Most previous attention models are based on low-level image features, as it is computationally and conceptually challenging to take into account highlevel factors such as scene context, topology or task. As a result, they often fail to predict saccadic targets because scene semantics strongly affect the planning and execution of fixations. In this talk, we present the first automated high level saliency predictor that incorporates the schema [Bartlett 1932] and singleton [Theeuwes and Godijn 2002] hypotheses into the Differential-Weighting Model (DWM) [Eckstein 1998]. The scene schema effect states that a scene is comprised of objects expected to be found in a specific context as well objects out of context which are salient (Figure 1a). The singleton effect refers to the finding that viewer’s attention is captured by isolated objects (Figure 1b). We propose a new model to account for high-level object saliency as predicted by the schema and singleton hypotheses by extending the DWM. The DWM models attentional processing using physiological noise in brain neurons and Gaussian combination rules. A GPU implementation of our model estimates the probabilities of individual objects to be foveated and is used in an innovative game level editor that automatically suggests game objects’ positioning. The difficulty of a game can then be implicitly adjusted since topology affects object search completion time.