The results reported in J. Wang, H. R. Tavakoli, and J. Laaksonen. Fixation Prediction in Videos using Unsupervised Hierarchical Features. CVPRW'17. were generated with compute_globalSaliency_video2img_onlymr.m.
One can find some pre-trained temporal ISA bases at ./bases, which were trained on random Youtube videos. Frame-by-frame images acquired from videos in ASCMN dataset (can be downloaded here) are required to make evaluation. Please unzip the downloaded .zip file, and place those files at ./videos, i.e.
./videos
+video1
-00001.jpg
-00002.jpg
-...
+video2
-00001.jpg
-....
+video2
...
...
+video24
An example to run compute_globalSaliency_video2img_onlymr.m is
compute_globalSaliency_video2img_onlymr(1, 7, 7, 0.1)
which evaluates video1 (the first variable) with temporal widths of the first and second ISA layers being 7 (the second and third variables), and sigma (the fourth variable) used in spatial prior being 0.1. Frame-by-frame results will be stored in ./results (We have provided some results there).
