This paper evaluates fixation prediction and salient object segmentation algorithms, along with statistics of major datasets. The authors identify serious design flaws in existing salient object benchmarks, called dataset design bias, due to overemphasis on stereotypical saliency concepts. This bias creates a disconnect between fixations and salient object segmentation and misleads algorithm design. A new high-quality dataset is proposed that includes both fixation and salient object segmentation ground-truth. By presenting fixations and salient objects simultaneously, the gap between them is bridged, and a novel salient object segmentation method is proposed. Significant benchmark progress is reported on three existing datasets. The paper discusses the relationship between fixation prediction and salient object segmentation, and proposes a new model combining fixation-based saliency models with segmentation techniques. This model significantly outperforms state-of-the-art algorithms on all three salient object datasets. The paper also discusses related works, including fixation prediction, salient object segmentation, objectness, object proposal, and foreground segments. It highlights the importance of dataset bias in benchmarking and the need for independent image acquisition and annotation to avoid bias. The authors analyze the PASCAL-S dataset, which includes both fixation and salient object annotations. They find that the segmentation maps of the Bruce dataset are significantly sparser than those of PASCAL-S or IS. They benchmark the F-measure of the test subset segmentation maps on the ground-truth masks and find that the performance of salient object segmentation algorithms drops significantly when migrating from the popular FT dataset. The paper also discusses the dataset design bias, which is caused by disproportionate sampling of positive/negative examples during dataset design. The authors propose a new model for salient object segmentation that combines existing techniques of segmentation and fixation-based saliency. This model outperforms previous methods by a large margin. The paper concludes that the problem of salient object segmentation should move beyond textbook examples of visual saliency and explore the strong correlation between fixations and salient objects. The proposed model decouples the problem into a segment generation process followed by a saliency scoring mechanism using fixation prediction. This simple model outperforms state-of-the-art algorithms on all major datasets.This paper evaluates fixation prediction and salient object segmentation algorithms, along with statistics of major datasets. The authors identify serious design flaws in existing salient object benchmarks, called dataset design bias, due to overemphasis on stereotypical saliency concepts. This bias creates a disconnect between fixations and salient object segmentation and misleads algorithm design. A new high-quality dataset is proposed that includes both fixation and salient object segmentation ground-truth. By presenting fixations and salient objects simultaneously, the gap between them is bridged, and a novel salient object segmentation method is proposed. Significant benchmark progress is reported on three existing datasets. The paper discusses the relationship between fixation prediction and salient object segmentation, and proposes a new model combining fixation-based saliency models with segmentation techniques. This model significantly outperforms state-of-the-art algorithms on all three salient object datasets. The paper also discusses related works, including fixation prediction, salient object segmentation, objectness, object proposal, and foreground segments. It highlights the importance of dataset bias in benchmarking and the need for independent image acquisition and annotation to avoid bias. The authors analyze the PASCAL-S dataset, which includes both fixation and salient object annotations. They find that the segmentation maps of the Bruce dataset are significantly sparser than those of PASCAL-S or IS. They benchmark the F-measure of the test subset segmentation maps on the ground-truth masks and find that the performance of salient object segmentation algorithms drops significantly when migrating from the popular FT dataset. The paper also discusses the dataset design bias, which is caused by disproportionate sampling of positive/negative examples during dataset design. The authors propose a new model for salient object segmentation that combines existing techniques of segmentation and fixation-based saliency. This model outperforms previous methods by a large margin. The paper concludes that the problem of salient object segmentation should move beyond textbook examples of visual saliency and explore the strong correlation between fixations and salient objects. The proposed model decouples the problem into a segment generation process followed by a saliency scoring mechanism using fixation prediction. This simple model outperforms state-of-the-art algorithms on all major datasets.