VATES: Video-Adaptive Temporal Enhancement System

Real-time video object recognition requires balancing accuracy and computational efficiency, yet existing approaches often compromise one for the other. Although modern detectors perform well on single images, their effectiveness on sequential data is limited by redundant inference, transient detection failures, and scale variation. To address these issues, we propose the Video-Adaptive Temporal Enhancement System (VATES), a model-agnostic inference pipeline that improves recognition quality without modifying the detector. VATES combines three lightweight components: keyframe-based skipping with linear interpolation, a short-term continuity mechanism for missed detections, and an adaptive input-scaling strategy driven by scene complexity. Together, these components enhance efficiency and robustness in diverse conditions. Experiments on 1080p surveillance video show that VATES significantly improves detection stability and inference speed, demonstrating its effectiveness as a general-purpose pipeline for video-based recognition.