OLBP

OLBP performs gaze-guided object segmentation by using personal fixation data to localize and produce boundary-preserving pixel-level segmentations of gazed objects.

Key Features:

• Novel Dataset Construction: The PFOS dataset was constructed by collecting pixel-level binary annotations over an existing fixation prediction dataset.
• Object Localization Module (OLM): An Object Localization Module interprets personal fixations to locate objects being gazed upon within images.
• Boundary Preservation Module (BPM): A Boundary Preservation Module introduces boundary information to maintain the completeness and integrity of segmented gazed objects.
• Mixed Bottom-Up and Top-Down Architecture: The network employs a mixed bottom-up and top-down approach that integrates multiple types of deep supervision.

Scientific Applications:

• Medical imaging: Enabling precise delineation of anatomical structures based on user gaze to support segmentation tasks.
• Human–computer interaction: Interpreting human gaze patterns to support intuitive gaze-driven interfaces and interaction-aware segmentation.

Methodology:

Constructed the PFOS dataset from pixel-level binary annotations over an existing fixation prediction dataset; used personal fixation data to guide segmentation by combining OLM and BPM within a mixed bottom-up and top-down network architecture integrating multiple types of deep supervision; evaluated against 17 state-of-the-art methods on PFOS.