Visual Saliency Prediction and Evaluation across Different Perceptual Tasks.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Cognitive Science

Background:

• Evaluating computational saliency models often relies on human eye-tracking fixation data.
• Fixation data presents challenges due to inherent biases, such as spatial bias, limiting evaluation validity.
• These biases stem from eye movement patterns independent of image content.

Purpose of the Study:

• To address limitations of fixation data in saliency model evaluation.
• To present modeling and evaluation of data from diverse perceptual tasks related to visual saliency.
• To introduce a novel benchmarking approach mitigating spatial bias issues.

Main Methods:

• Developed and evaluated models using data from various perceptual tasks.
• Introduced a novel benchmarking methodology to counter spatial bias in eye-tracking data.
• Demonstrated an approach to approximate alternative perceptual task outputs using computational saliency and eye gaze data.

Main Results:

• Established the value of alternative perceptual data over fixation data for model improvement.
• Presented novel benchmarking results and methods for visual saliency research.
• Achieved a new performance baseline for perceptual tasks offering insights into visual saliency.

Conclusions:

• Alternative perceptual tasks provide a valuable window into visual saliency, complementing or replacing fixation data.
• The proposed benchmarking approach effectively addresses spatial bias challenges.
• Computational saliency can approximate human behavior in specific visual tasks, enhancing model applicability.