Visual Attention Modeling for Stereoscopic Video: A Benchmark and Computational Model.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Multimedia Signal Processing

Background:

• Visual attention modeling is crucial for understanding human perception in stereoscopic video.
• Existing models often lack accuracy in predicting attention for 3D content.
• A robust benchmark database is needed for evaluating stereoscopic visual attention models.

Purpose of the Study:

• To develop a novel computational model for visual attention in stereoscopic video.
• To create a large-scale eye-tracking database for benchmarking stereoscopic attention models.
• To improve the accuracy of saliency prediction in 3D videos.

Main Methods:

• Construction of a large-scale eye-tracking database with 47 stereoscopic video sequences and fixation data.
• Development of a computational model based on Gestalt theory, extracting luminance, color, texture, and depth features.
• Calculation of spatial saliency using feature contrast and temporal saliency using motion contrast (planar and depth).
• Fusion of spatial and temporal saliency with uncertainty weighting derived from Gestalt principles (proximity, continuity, common fate).

Main Results:

• The proposed model demonstrates superior performance compared to state-of-the-art methods.
• The model achieves high accuracy on the newly built large-scale eye-tracking database.
• Validation of the model's effectiveness on an additional 3D video database (DML-ITRACK-3D).

Conclusions:

• The developed computational model effectively predicts visual attention in stereoscopic videos.
• The created eye-tracking database serves as a valuable benchmark for future research.
• The Gestalt theory-based approach offers a promising direction for advanced saliency detection in 3D media.