Please note that the new CAT2000 evaluation is not yet finished. Results might be subject to small changes.

Click on a metric name to sort that metric. Models evaluated as probabilistic models are shown with green background. The performance under the metric which a model has been trained on is shown in italic. The code for evaluation models can be found here.

Name | Published | Code | IG | AUC | sAUC | NSS | CC | KLDiv | SIM | Date tested |

Lingyun Zhang, Matthew H. Tong, Tim K. Marks, Honghao Shan, Garrison W. Cottrell. SUN: A Bayesian framework for saliency using natural statistics [JoV 2008] |
0.7128 | 0.5815 | 0.8127 | 0.2902 | 2.3353 | 0.3982 |
First tested 2021-04-07 Last tested 2021-04-07 maps from code via SMILER. Params: rescale=0.5 |
|||

Saliency Detection by Self-Resemblance (SSR) |
Hae Jong Seo, Peyman Milanfar Nonparametric Bottom-Up Saliency Detection by Self-Resemblance [CVPR 2009]. |
0.7209 | 0.5943 | 0.8543 | 0.3047 | 1.7459 | 0.4110 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

Quaternion-Based Spectral Saliency (QSS) |
B. Schauerte, R. Stiefelhagen. Quaternion-based Spectral Saliency Detection for Eye Fixation Prediction [ECCV 2012] |
0.7510 | 0.6132 | 0.8368 | 0.3015 | 1.2136 | 0.4065 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

AIM |
Neil Bruce, John Tsotsos. Attention based on information maximization [JoV 2007] |
0.7547 | 0.6092 | 0.9680 | 0.3479 | 1.3744 | 0.4093 |
First tested 2021-04-07 Last tested 2021-04-07 maps from code via SMILER. Params: resize=0.5, thebasis='31infomax975' |
||

OpenSALICON |
Christopher Lee Thomas. OpenSalicon: An Open Source Implementation of the Salicon Saliency Model [arXiv 2016] |
0.7627 | 0.6341 | 1.1985 | 0.4142 | 1.1731 | 0.4573 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

Dynamic Visual Attention (DVA) |
Hou, Xiaodi, and Liqing Zhang. Dynamic visual attention: Searching for coding length increments NIPS 2008 |
0.7664 | 0.6089 | 1.0686 | 0.3778 | 1.3032 | 0.4438 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
||

Nicolas Riche, Matei Mancas, Matthieu Duvinage, Makiese Mibulumukini, Bernard Gosselin, Thierry Dutoit. RARE2012: A multi-scale rarity-based saliency detection with its comparative statistical analysis [Signal Processing: Image Communication, 2013] |
0.7689 | 0.6108 | 1.0960 | 0.3914 | 1.1489 | 0.4382 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
|||

Image Signature |
Xiaodi Hou, Jonathan Harel, Christof Koch. Image Signature: Highlighting Sparse Salient Regions [PAMI 2011] |
0.7704 | 0.6082 | 1.0864 | 0.3898 | 1.1073 | 0.4283 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

Stas Goferman, Lihi Zelnik-Manor, Ayellet Tal. Context-Aware Saliency Detection [CVPR 2010] [PAMI 2012] |
0.7775 | 0.5939 | 1.0901 | 0.3934 | 1.1093 | 0.4408 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
|||

Jianming Zhang, Stan Sclaroff. Saliency detection: a boolean map approach [ICCV 2013] |
0.7826 | 0.6267 | 1.1879 | 0.4190 | 1.2060 | 0.4615 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
|||

ML-Net |
Marcella Cornia, Lorenzo Baraldi, Giuseppe Serra, Rita Cucchiara. A Deep Multi-Level Network for Saliency Prediction [ICPR 2016] |
0.7852 | 0.6318 | 1.3280 | 0.4631 | 1.1078 | 0.4774 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

IttiKoch2 |
Implementation by Jonathan Harel (part of GBVS toolbox) |
0.7916 | 0.5954 | 1.1945 | 0.4318 | 1.0269 | 0.4505 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

Jonathan Harel, Christof Koch, Pietro Perona. Graph-Based Visual Saliency [NIPS 2006] |
0.7979 | 0.5863 | 1.2342 | 0.4460 | 1.0169 | 0.4594 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
|||

Deep Visual Attention (DVA) |
W. Wang, and J. Shen. Deep Visual Attention Prediction [IEEE TIP 2018] |
0.8000 | 0.6253 | 1.4287 | 0.5033 | 0.9931 | 0.4924 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
||

Junting Pan, Cristian Canton, Kevin McGuinness, Noel E. O'Connor, Jordi Torres, Elisa Sayrol and Xavier Giro-i-Nieto. SalGAN: Visual Saliency Prediction with Generative Adversarial Networks [arXiv 2017] |
0.8085 | 0.6354 | 1.4624 | 0.5207 | 1.1155 | 0.4921 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
|||

Hamed Rezazadegan Tavakoli, Esa Rahtu, Janne Heikkila. Fast and efficient saliency detection using sparse sampling and kernel density estimation [SCIA 2011] |
0.8212 | 0.5450 | 1.6103 | 0.5799 | 2.6123 | 0.5255 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
|||

LDS |
Shu Fang, Jia Li, Yonghong Tian, Tiejun Huang, Xiaowu Chen. Learning Discriminative Subspaces on Random Contrasts for Image Saliency Analysis [TNNLS 2016] |
0.8281 | 0.5669 | 1.5356 | 0.5574 | 1.0453 | 0.5267 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |
||

Erkut Erdem, Aykut Erdem. Visual saliency estimation by nonlinearly integrating features using region covariances [JoV 2013] |
0.8402 | 0.5570 | 1.7411 | 0.6251 | 1.5630 | 0.5603 |
First tested 2021-03-31 Last tested 2021-03-31 maps from SMILER |
|||

Eleonora Vig, Michael Dorr, David Cox. Large-Scale Optimization of Hierarchical Features for Saliency Prediction in Natural Images [CVPR 2014] |
0.8470 | 0.5782 | 1.2092 | 0.4482 | 1.1856 | 0.3997 |
First tested 2021-04-01 Last tested 2021-04-01 maps from SMILER |

- Gold Standard: Our gold standard model is a Gaussian Kernel Density Estimate. There are two versions of it. The crossvalidated performance is the leave-one-subject-out performance where for each subject and image the fixations of all other subject on the same image are used to construct a kernel density estimate that is then evaluated on the remaining subject. The kernel size and the mixture weight of a uniform regularization component are fitted by maximizing the cross-validated log-likelihood of the model. In addition to this crossvalidated version of the model, we also report the performance of a KDE model that uses all fixations on each image with the same parameters as the cross-validated model. One can interpret the cross-validated performance as a lower bound on the explainable performance and the joint performance as an upper bound.
- Centerbias: The center bias model is again a Gaussian Kernel Density Estimate. However, unlike the gold standard, it uses the fixations of all
*other*images to predict the fixations on any given image. Kernel size and the mixture weight oa a uniform regularization component are again fitted by maximizing the model log-likelihood.