Deep Networks for Saliency Detection via Local Estimation and Global Search Supplementary Material

We evaluate our method (LEGS) on four benchmark data sets (MSRA-5000 [7], SOD [9], ECCSD [10] and PASCALS [6]) against ten state-of-the-art models (SVO [1], PCA [8], DRFI [3], GC [2], HS [10], MR [11], UFO [4], wCtr [12], CPMC-GBVS [6] and HDCT [5]). We use either the implementations or the saliency maps provided by the authors for fair comparison. Since the DRFI method is also trained on the MSRA-5000 data set with different training/test images from ours, we do not report its result on this data set. In addition, the CPMC-GBVS method only provides the saliency maps of the PASCAL-S data set. Both quantitative and qualitative results are demonstrated in the following figures.