Abstract
When viewing natural images, primates fixate locations of interest. In particular, both bottom-up saliency (e.g. contrast) and top-down target relevance influence where to look next. Computational models that take both saliency and relevance into account, predict fixations better than those based on saliency alone. While previous studies have analyzed the effect of relevance, in these studies both feature-based and location-based attention could have influenced search behavior. Here, to focus on feature-based attention alone, we designed a search task that does not have any spatial cues; two macaque monkeys searched for a 2° x 2° vertical Gabor target placed randomly (uniformly) within human-photographed scenes. The data comprised ~11,000 fixations spanning ~1600 different image searches. We then asked (1) how well relevance predicts gaze and (2) what features the fixated image patches shared with the target. First, to quantify how well relevance predicts gaze, we defined relevance of a fixated image patch as the maximum of the convolution of the fixated image patch with the target Gabor wavelet. We found that across all fixations (excluding the last), the monkey fixated at locations that were more relevant than predicted by chance (Area under ROC, or AUC=0.66 and 0.65 vs. shuffle predictor, for monkeys 1 and 2). Furthermore, relevance predicted fixations far better than the Itti-Koch saliency model (AUC=0.51 and 0.50, p<0.001). This result suggests that monkeys preferentially fixate task-relevant locations, and this behavior supersedes the effect of other salient locations. Second, we quantified the similarity between target and fixated image patches by examining their spectral properties. We found that the fixated patches have higher energy than patches of shuffled images, in particular, along the relevant orientation. This result shows that monkeys tend to fixate at locations sharing spectral properties with the target.
Meeting abstract presented at VSS 2013