What is a salient object? A dataset and a baseline model for salient object detection.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Salient object detection models aim to identify and segment important objects in images.
• Existing datasets often lack multiple objects, potentially biasing model evaluation.
• The relationship between human visual attention (fixations) and perceived saliency is not fully understood.

Purpose of the Study:

• To investigate the correlation between human fixations and explicit saliency judgments.
• To introduce new, less biased benchmark datasets for evaluating salient object detection models.
• To propose a robust baseline model for comparison and identify limitations in current approaches.

Main Methods:

• Analyzed the relationship between human eye fixations and explicit saliency judgments on scene datasets.
• Developed two novel benchmark datasets featuring scenes with multiple objects.
• Proposed a superpixel-based model as a baseline for salient object detection.

Main Results:

• Established that the most salient object attracts the highest fraction of human fixations.
• Observed a significant performance drop (40%-70%) in state-of-the-art models on the new datasets.
• The proposed baseline model achieved competitive results, outperforming others on new data and revealing segmentation/localization conflation in existing models.

Conclusions:

• Human fixation patterns provide a strong indicator for salient object identification.
• Current salient object detection models may be over-fitted to biased datasets.
• The proposed datasets and baseline model offer a more rigorous evaluation framework for future research.