Centre-Negative: An Effective and Efficient Solution to Centre Bias in Visual Saliency Evaluation
Keywords: saliency prediction, centre bias, metric
TL;DR: We propose a new saliency metric which can evaluate centre bias on the boarder case that previous metrics will fail, and our solution is more efficient.
Abstract: Spatial bias is a long-standing problem in visual saliency detection, and various evaluation metrics have been proposed to address this issue. In this paper, we first review existing bias-specific saliency metrics, we group them into fixation-based and region-based and show that all of the metrics suffer from different drawbacks, especially when most fixations are closely distributed near the centre as is typical. To solve this problem, we study the essence of spatial bias in visual saliency. We show that the bias could be important in modelling saliency signal, thus ignoring or penalizing central regions cannot measure a saliency model comprehensively. The bias becomes problematic when comparing different saliency algorithms, a model with a strong central preference can capture most fixations. From this perspective, we propose a region-based metric-agnostic solution called Centre-Negative. The proposed approach can deliver three main advantages: a) our solution is not designed for a saliency metric specifically, Centre-Negative can be combined with any existing metrics to make use of their properties and simultaneously overcome the centre bias problem; b) our method is a region-based solution, so it can handle the situation where most fixations are densely distributed near the centre; c) our method can be applied simply and efficiently with a time complexity of $\mathcal{O}(1)$, and we show that the negative map created by Centre-Negative outperforms other solutions.
Submission Type: Full Paper
Submission Number: 6
Loading