Elizabeth Y. Zhou¹, Yelda Semizer², Melchi M. Michel¹; ¹Rutgers University, ²New Jersey Institute of Technology

Vision researchers have long studied the effect of clutter on visual search performance. We are curious about the influence of the observer's task on image clutter. In Semizer and Michel (2022), participants searched for objects in natural images after they were shown the category name of the objects. In this case, the “category-level” metric for target-related clutter was computed based on the discrepancy between the category representation and the image representation (i.e., the distance between the feature distribution across all exemplars of the target category within the image set and the feature distribution of the image being searched). Zhou et al. (2024) showed that the category-level metric was able to predict search time either when a target is present or absent in the image. Here, we explored whether the category representation is invariant across orientations. When building the feature distribution for the category representation, we rotated each exemplar to eight random orientations within some selected ranges. We observed that as the possible range of rotation increases (30, 60, 90, 180 and 360 degrees), the performance of our category-level metric decreases in predicting the search time. Also, we investigated the effects of richness in the category representation. As the dimensionality of feature increases when building the category representation (12, 50, 100 and 200 components from principal component analysis), the performance of our category-level metric decreases in predicting the search time when the search target is absent, while the trend is mixed when the search target is present. These results potentially reflect limited orientation invariance and limited feature resolution in human category representations that guide visual search.