A Dual-Branch Visual-Textual Model for Contextual Scene Awareness

Project Overview

Conventional captioners name objects but not the relationships between them. This model reasons over structure before it writes a word using panoptic segmentation for pixel-level scene parsing and scene-graph decoding for relational understanding. The dual-branch architecture simultaneously processes visual features and textual semantics for contextually rich captions.