A Behavioral Analysis of Computational Models of Visual Attention

Abstract

Abstract. Robots often incorporate computational models of visual attention to streamline processing. Even though the number of visual attention systems employed on robots has increased dramatically in recent years, the evaluation of these systems has remained primarily qualitative and subjective. We introduce quantitative methods for evaluating computational models of visual attention by direct comparison with gaze trajectories acquired from humans. In particular, we focus on the need for metrics based not on distances within the image plane, but that instead operate at the level of underlying features. We present a framework, based on dimensionality-reduction over the features of human gaze trajectories, that can simultaneously be used for both optimizing a particular computational model of visual attention and for evaluating its performance in terms of similarity to human behavior. We use this framework to evaluate the Itti et al. (1998) model of visual attention, a computational model that serves as the basis for many robotic visual attention systems.

Multimodal saliency-based attention for object-based scene analysis

“…We perform a couple of experiments to evaluate the behavior of the proposed system, because a quantitative, comparative method to evaluate the performance of an overt attention system does not exist (see [9], [17]). In order to obtain a reliable impression of the performance of our system, we repeated every experiment multiple times with varying environmental conditions such as, e.g., lighting, number of objects, distracting clutter, and timing of events.…”

Multimodal saliency-based attention for object-based scene analysis

“…For this purpose, we introduce a novel definition of auditory attention based on Bayesian surprise [16]. In contrast, in computer vision a huge amount of saliency models has been proposed in recent years (see [11], [17]). Since reviewing them is beyond the scope of this paper, we recommend reading the survey of computational visual attention in [11].…”

From Robot Arm to Intentional Agent: the Articulated Head

“…A substantial amount of knowledge has been accumulated, though sometimes disparate or conflicting. One of the most important findings for attention systems in machines (Shic & Scassellati, 2007) is that attention is driven by two sources, saliency in the perceptual input (bottom-up or exogenous attention) and task-dependent attention direction (top-down or endogenous attention). The former is comparatively easier to handle and evaluate (e.g.…”