Scholarlink

<p>As we explore and interact with the real world, we need to decide where to focus our attention next. The outcome of this decision influences our responses to various actions in the environment. Also scanpath prediction enhances implicit cues for task relevance and spatial awareness, optimizing the ability of LLMs and vision models to interpret visual data by modeling human-like gaze dynamics and emphasizing salient regions. The present research work analyzes features extracted and found that a combination of various parameters such as depth, presence of humans, objects, text can help predict scanpaths in real-life scenarios. A deep learning model incorporated with encoder part of transformer is utilized to find a priority map from RGB and depth images that depicts how the human fixations can happen while viewing a natural image with multiple objects present in the scene and the evaluation metrics indicates the model got a superior performance. The experiment examines the directions of scanpaths while viewing humans and objects, as well as the contributions of color, motion, and depth. These findings unearth the human perspectives while viewing a scene which can be applied to robotic navigation, human interests, and other areas.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling human attention: analyzing scanpaths and visual features in fixation prediction with transformer-based deep learning

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Judy K. George,
  • Elizabeth Sherly

摘要

As we explore and interact with the real world, we need to decide where to focus our attention next. The outcome of this decision influences our responses to various actions in the environment. Also scanpath prediction enhances implicit cues for task relevance and spatial awareness, optimizing the ability of LLMs and vision models to interpret visual data by modeling human-like gaze dynamics and emphasizing salient regions. The present research work analyzes features extracted and found that a combination of various parameters such as depth, presence of humans, objects, text can help predict scanpaths in real-life scenarios. A deep learning model incorporated with encoder part of transformer is utilized to find a priority map from RGB and depth images that depicts how the human fixations can happen while viewing a natural image with multiple objects present in the scene and the evaluation metrics indicates the model got a superior performance. The experiment examines the directions of scanpaths while viewing humans and objects, as well as the contributions of color, motion, and depth. These findings unearth the human perspectives while viewing a scene which can be applied to robotic navigation, human interests, and other areas.