Improving saliency models' predictions of the next fixation with humans' intrinsic cost of gaze shifts

TL;DR

This work addresses the gap in predicting human gaze targets by integrating intrinsic gaze costs into saliency-based predictions through a sequential decision framework. It converts static saliency maps into dynamic, history-aware value maps by combining a saliency-based reward $r_{ ext{free view}}=S(I,\boldsymbol{x_{t+1}})$ with an independently measured oculomotor cost map $r_{ ext{internal}}$ and a history-driven exploration bonus $r_{ ext{fixation history}}$, merged as $r(s,\boldsymbol{x_{t+1}})= w_0 r_{ ext{free view}} + w_1 r_{ ext{internal}} + w_2 r_{ ext{fixation history}}$ with $w_0=1$. The model estimates parameters $\phi_i$, $\boldsymbol{\Sigma}$, $w_1$, and $w_2$ from human data and uses a greedy one-step look-ahead to compute a sequence of value maps, significantly improving next-fixation NSS and AUC across five strong saliency models on MIT1003, OSIE, and Toronto datasets. Key findings show that past fixations gain value over time and that eye-movement costs interact with image content to shape gaze targets, challenging the idea that gaze sequences are independent of behavioral preferences. The approach provides a general, plug-in method to enhance arbitrary saliency models by incorporating intrinsic gaze costs, with practical impact for gaze prediction, human-computer interaction, and visual scene understanding.

Abstract

The human prioritization of image regions can be modeled in a time invariant fashion with saliency maps or sequentially with scanpath models. However, while both types of models have steadily improved on several benchmarks and datasets, there is still a considerable gap in predicting human gaze. Here, we leverage two recent developments to reduce this gap: theoretical analyses establishing a principled framework for predicting the next gaze target and the empirical measurement of the human cost for gaze switches independently of image content. We introduce an algorithm in the framework of sequential decision making, which converts any static saliency map into a sequence of dynamic history-dependent value maps, which are recomputed after each gaze shift. These maps are based on 1) a saliency map provided by an arbitrary saliency model, 2) the recently measured human cost function quantifying preferences in magnitude and direction of eye movements, and 3) a sequential exploration bonus, which changes with each subsequent gaze shift. The parameters of the spatial extent and temporal decay of this exploration bonus are estimated from human gaze data. The relative contributions of these three components were optimized on the MIT1003 dataset for the NSS score and are sufficient to significantly outperform predictions of the next gaze target on NSS and AUC scores for five state of the art saliency models on three image data sets. Thus, we provide an implementation of human gaze preferences, which can be used to improve arbitrary saliency models' predictions of humans' next gaze targets.

Figures (7)

• Figure 1: Schematic of the algorithm. An arbitrary saliency map and the scanpath with the current gaze position are the input. Output is a value map, which integrates the saliency map, the recomputed map for the cost of gaze shifts, and the sequential history dependant map. Note that the original image is not an input to the algorithm.
• Figure 2: Example predictions of the next fixation. Each row shows the original image together with the respective preceding scanpath together with the current $i$-th fixation marked with a cross. The corresponding saliency, cost, and exploration maps as well as the final value map are shown from left to right. The predicted fixation is shown together with the ground truth next fixation of the human observer marked with a diamond.
• Figure 3: Estimated exploration values for four different underlying saliency models (blue, orange, red, green) and the corresponding averaged curve (black). Note that the estimated $\phi_i$ from Table S1 have been multiplied here by their associated weight $w_2$ to visualize the actual influence of the exploration map.
• Figure 4: Differences in the NSS scores between our dynamic value maps and the underlying static saliency maps. Positive values indicate that our dynamic model predicted the subsequent fixation better than the baseline model. The errorbars indicate $\pm$ standard error of the mean.
• Figure S1: NSS scores for the one-step ahead prediction depending on the ordinal position in the gaze sequence.