**4.2 Results**

## *4.2.1 First experiment: Multiple networks*

In order to evaluate how accurately the saliency map is able to match the location of human fixations, it used a set of metrics previously defined by [17].

In **Table 4** we show results of area under ROC (AUC), correlation coefficient (CC), normalized scanpath saliency (NSS), Kullback-Leibler divergence (KL), and similarity (SIM) for every network for all datasets.

The area under ROC (AUC) is considered as true positives, the saliency map values coincide with a fixation and false positives, and the saliency map values that have no fixation then compute the area under the curve. Similarly, the NSS computes the


#### **Table 1.**

*Characteristics of eye-tracking datasets.*

*Saliency Detection from Subitizing Processing DOI: http://dx.doi.org/10.5772/intechopen.108552*


#### **Table 2.**

*Comparison of our saliency output with standard benchmark methods over real image Toronto dataset for saliency prediction. (Top) Baseline low-level saliency models. (Bottom) State-of-the-art deep saliency models. Best score for each metric is defined as bold and TOP-3 scores are italicized.*


#### **Table 3.**

*Comparison of our saliency output with standard benchmark methods over synthetic image SID4VAM dataset for saliency prediction. (Top) Baseline low-level saliency models. (Bottom) State-of-the-art deep saliency models. Best score for each metric is defined as bold and TOP-3 scores are italicized.*


#### **Table 4.**

*Benchmark of our method with different networks (top 1 networks are italicized).*

average normalized saliency map that coincides with fixations. Other metrics such as CC, KL, and SIM compute the score upon the region distribution statistics of all pixels (KL calculating the divergence and CC/SIM the histogram intersection or similarity of the distribution).

After computing the saliency maps for all datasets (see in **Table 4**) with AlexNet and ResNet152, we observed that metric scores vary considerately depending on dataset or network. AlexNet is shown to provide best results for pop-out patterns (SID4VAM), whereas ResNet152 shows overall higher scores with real images of Toronto dataset.
