Introducing context-dependent and spatially-variant viewing biases in saccadic models

Abstract

Previous research showed the existence of systematic tendencies in viewing behavior during scene exploration. For instance, saccades are known to follow a positively skewed, long-tailed distribution, and to be more frequently initiated in the horizontal or vertical directions. In this study, we hypothesize that these viewing biases are not universal, but are modulated by the semantic visual category of the stimulus. We show that the joint distribution of saccade amplitudes and orientations significantly varies from one visual category to another. These joint distributions are in addition spatially variant within the scene frame. We demonstrate that a saliency model based on this better understanding of viewing behavioral biases and blind to any visual information outperforms well-established saliency models. We also propose a saccadic model that takes into account classical low-level features and spatially-variant and context-dependent viewing biases. This model outperforms state-of-the-art saliency models, and provides scanpaths in close agreement with human behavior. The better description of viewing biases will not only improve current models of visual attention but could also influence many other applications such as the design of human-computer interfaces, patient diagnosis or image/video processing applications.

Problems with Saliency Maps

“…In such perspective, Le Meur and colleagues [26] have proposed saccadic models as a new framework to predict visual scanpaths of observers while they freely watch static images. In such models the visual fixations are inferred from bottom-up saliency and oculomotor biases (captured as saccade amplitudes and saccade orientations) that are modeled using eye tracking data.…”

“…Performance of these models can be evaluated either by directly comparing the generated scanpaths to human scanpaths or by computing new saliency maps, in the shape of densities from model generated fixations. There is a limited number of saccadic models available, see [26] for a comprehensive review; generalisation to dynamic scenes have been presented for instance in [4,28]. A remarkable result obtained by saccadic models is that by using simulated fixations { r (2), · · · } to generate a model-based fixation map, the latter has higher predictive performance than the raw salience map S, in terms of similarity/dissimilarity µ with respect to human fixation maps.…”

Advanced Statistical Methods for Eye Movement Analysis and Modelling: A Gentle Introduction

“…To summarise, by simply taking into account the prior P (x), a richness of possible behaviours and analyses are brought into the game. To further explore this perspective, we recommend the thorough and up-to-date review by Le Meur and Coutrot [90].…”

Perception-driven Accelerated Rendering

“…Apparently, when performing a saccade, humans are biased towards making horizontal or vertical saccades. Le Meur et al [LMC16,OLME16] exploit this bias in combination with bottom-up feature detection, resulting in a saccadic model for freeviewing scenarios. The authors make use of their model in a method for spatial fixation prediction and scanpath generation.…”