Table 2.
Quantitative metrics of saliency prediction on the test set. “SS-att” indicates those single-stream models for saliency prediction without a classification branch. Saliency metrics include information gain (IG), pearson’s cross-correlation (CC), normalized saliency scan path (NSS), similarity (SIM), and area under curve (AUC) [16].
| Models | IG | CC | NSS | SIM | AUC |
|---|---|---|---|---|---|
| M-SEN BCE + GAN | 0.543 | 0.693 | 2.525 | 0.512 | 0.775 |
| M-SEN BCE | 0.429 | 0.615 | 2.144 | 0.469 | 0.726 |
| M-SEN MSE + GAN | 0.307 | 0.634 | 2.327 | 0.309 | 0.616 |
| M-SEN MSE | 0.288 | 0.556 | 2.253 | 0.310 | 0.603 |
| SS-att BCE | 0.192 | 0.708 | 1.480 | 0.570 | 0.801 |
| SS-att MSE | 0.152 | 0.546 | 1.329 | 0.532 | 0.788 |