Abstract
We propose GTANet – a novel approach to learning the alignments between human gaze scanpaths and fine-grained task descriptions in vision-language tasks. While the influence of tasks on gaze is well known, the relationship between gaze scanpaths and fine-grained task descriptions remains largely unexplored. GTANet addresses this gap by aligning encoded spatiotemporal gaze features with text descriptions. We utilize a patch-based gaze encoder to generate gaze features that reflect visual contexts, and a multimodal feature mixer to fuse the gaze features and the task descriptions, capturing cross-modal alignment. To validate our method, we introduce two novel tasks: gaze-to-question and question-to-gaze retrieval. Experiments on the AiR and MHUG datasets demonstrate that GTANet consistently outperforms baseline methods across all Recall@K metrics, achieving substantial improvements in both retrieval directions. These results confirm the strong link between human gaze and fine-grained task descriptions, thus validating the effectiveness of our approach.