scanpath

scanpath predicts human visual scanpaths to model overt human visual attention and search behavior during task-free viewing.

Key Features:

• IOR-ROI Recurrent Mixture Density Network Framework: A framework based on Inhibition of Return (IOR) and Region of Interest (ROI) recurrent mixture density networks that predicts ordered fixation positions and their durations.
• Integration of Features: Integrates bottom-up features from raw visual stimuli and semantic features extracted by convolutional neural networks as inputs to the IOR-ROI LSTM network.
• Dual LSTM Units: Uses an IOR-LSTM that adaptively maintains and updates information about previously fixated regions to capture inhibition of return dynamics and an ROI-LSTM that predicts next ROIs using spatially inhibited image feature maps on a feature-wise basis.
• Fixation Duration Prediction: Employs a regression neural network to predict fixation durations from viewing history and current ROI image features.
• Mixture Density Network for Variability: Models next-fixation location distributions as Gaussian mixtures and fixation duration distributions as Gaussians via a mixture density network to capture inter-subject variability.

Scientific Applications:

• Modeling overt visual attention: Predicts sequential eye-movement trajectories and fixation patterns during task-free viewing to study gaze allocation.
• Cognitive neuroscience: Supports investigations linking fixation sequences and durations to perceptual and cognitive processes.
• Psychology: Assists research on visual search, attention distribution, and gaze-related behavioral measures.

Methodology:

Train the IOR-ROI recurrent mixture density network on the OSIE (Oxford Student Eye-tracking) and MIT low-resolution eye-tracking datasets and evaluate against these benchmarks; the approach uses bottom-up and CNN-extracted semantic feature maps, IOR-LSTM and ROI-LSTM units, a regression neural network for fixation durations, and a mixture density network modeling Gaussian mixtures for next-fixation locations and Gaussian distributions for durations.