Video Frame Selection

Video frame selection studies how to choose a small set of frames that preserves the events, objects, and temporal context needed for a task while reducing computation, storage, or annotation. It matters in long-video question answering, surveillance, robotics, medical imaging, and compression, where processing every frame is expensive but missing a brief event can change the result. Methods range from uniform or shot-based sampling and motion or novelty filters to graph- and coverage-based diversity, query-conditioned relevance ranking, and adaptive selectors that use uncertainty, attention, or event structure; selected frames may also receive different resolutions or budgets. Current work increasingly integrates selection with video-language models and edge systems, but a central tension remains: task-specific evidence can improve accuracy while making selections less reusable, and sparse samples risk losing event order, transitions, or rare events, making careful benchmark and budget-matched comparisons important.