Active visual perception lets an intelligent system choose how it gathers visual evidence rather than passively processing a fixed image. A review posted on arXiv in December 2025 (and revised in October 2026) attempts to bring order to the field by separating the act of acquiring observations from the processing of already-captured inputs. The authors organize the literature around four questions: what evidence a task requires, which actions can obtain that evidence, how candidate observations should be valued, and how the next sensing action should be selected and executed.
To judge whether active perception actually helps, the review examines datasets, interactive environments, physical platforms, and baseline comparisons. It proposes an evaluation framework that links improvements in visual evidence to downstream task outcomes while also accounting for acquisition costs and risks. The authors argue that two common success indicators are misleading: predicted observation value does not guarantee that the evidence can be acquired through feasible actions, and task success alone does not prove that active observation selection was the reason.
Going forward, the review identifies open challenges: reliable value estimation, planning over longer decision sequences under physical and resource constraints, grounding semantic reasoning in acquired evidence, and maintaining collaboration under imperfect communication. It emphasizes that progress should be verified through reproducible evaluation of the complete perception-and-action loop.