Abstract

Visual similarity relationships underlie a host of human behaviors, and determining these relationships is crucial both to the understanding of these behaviors and the construction of automated systems designed for human use. We conducted a large-scale web-based experiment in which subjects rank ordered random objects according to their visual similarity to object classes. We then constructed a computational model, using Adaboost with color, texture, and shape features, to quantify the visual similarity between these objects and the target classes, and to perform the same similarity ranking task. The model and subjects showed good agreement in the objects judged to be most and least similar to the target categories. Our data also suggest that color, texture, and shape features are all useful for classification, and that the specific weighting of these features depends on the target object class. Moreover, we show that these target-like properties constitute a learnable class, as it was possible to train a classifier on target-like objects (i.e., without positive samples) to recognize actual targets.