Learning to Predict Sequences of Human Visual Fixations.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Cognitive Science

Background:

• Current visual attention models generate saliency maps but do not predict the temporal sequence of eye fixations.
• Predicting fixation sequences is crucial for understanding visual exploration and improving attention models.

Purpose of the Study:

• To develop a method for predicting the temporal sequence of human eye fixations.
• To learn a visual exploration policy that mimics human eye movement patterns.

Main Methods:

• Utilized least-squares policy iteration (LSPI) to learn a visual exploration policy from recorded human eye-tracking data.
• Developed a model with stage-specific parameters to capture cue importance throughout visual exploration (scanpath).

Main Results:

• Demonstrated the effectiveness of LSPI in combining multiple cues at different stages of the scanpath.
• Found that low-level and high-level (semantic) cues are equally important for the first fixation, with semantic cues increasing in importance over time.
• Achieved state-of-the-art performance on the OSIE and MIT datasets.

Conclusions:

• The proposed method effectively predicts human eye fixation sequences.
• The findings provide insights into the dynamic interplay of visual cues during human visual exploration.
• The approach advances the field of computational visual attention by incorporating temporal dynamics.