Active Visual Perception
Overview
Active visual perception studies how an agent’s actions—such as shifting gaze, zooming into a crop, moving a camera, or changing illumination—can improve what it sees for a task. Methods include biologically inspired foveation and scanpaths, adaptive image and token selection, and learned camera or robot policies; multimodal language models increasingly use these tools to gather evidence for video, document, remote-sensing, and 3D reasoning. Evaluation compares active observation with fixed or full-image input, using measures such as task accuracy, manipulation success, scanpath similarity, and sensing cost. Results show that selective viewing can save computation and improve performance, but benefits depend on the task, and a key challenge is choosing observations that are both informative and feasible while evaluating them without misleading shortcuts such as center bias.
Latest papers
49 papers9 papers in the last four weeks, down 10% on the four weeks before. 0.1% of all new papers.