Deep Contrast Learning for Salient Object Detection (Supplemental Materials)

PASCAL-S is a dataset built using the validation set of the PASCAL VOC 2010 segmentation challenge. It contains 850 images with the ground truth labeled by 12 subjects. We threshold the masks at 0.5 to obtain binary masks as suggested in [4]. The SOD dataset contains 300 images and it was originally designed for image segmentation. Many images in this dataset have multiple salient objects with low contrast. We compare our saliency models (DCL and DCL) against eight recent state-of-the-art methods, including SF [6], GC [1], DRFI [2], PISA [8], BSCA [7], LEGS [9], MC [10] and MDF [3]. The last three are the latest deep learning based methods. For fair comparison, we use either the implementations or the saliency maps provided by the authors. In addition, we also train a fully convolutional neural network (FCN) (the FCN-8s network proposed in [5]) for comparison. As shown in Fig. 1, our method achieves the highest PR curve on both PASCAL-S and SOD datasets. The average precision, recall and F-measure on these two datasets is shown in Fig. 2, sorted by F-measure. Our method achieves the best performance on the overall F-measure as well as significant improvements in both precision and recall. On the PASCAL-S dataset, our DCL with CRF (DCL) achieves 78.55% precision and 77.07% recall while the second best (MDF [3]) achieves 81.14% precision and 61.75% recall. On the SOD dataset, our method achieves 85.67% precision and 64.75% recall while the second best (MDF[3]) achieves 85.72% precision and only 58.32% recall. Although the precision of MDF is higher than ours, its recall is significantly lower. Thus it is much more likely for MDF to miss salient objects. This is also reflected in the lower F-measure and higher MAE achieved with MDF.