Piecewise Video Condensation for Complex Scenes

Abstract

Video synopsis or condensation provides an efficient way to video storage and browsing. Lots of improvements have been made for boosting the speed or improving the condensed quality, which have shown promising results. However, most of the existing approaches cannot effectively deal with complex scenes, such as sudden changes or background object movement. In this paper, we propose a robust video condensation approach for complex scenes. A video segmentation method is designed to analyze the background complexity and divide the input video into several segments. The advantage is two-fold: one is to judge the complexity of backgrounds; the other is to generate a piecewise background image for each segment. Then, we adopt a divide-and-conquer strategy for video condensation. We keep the original video segments for complex backgrounds while maximally condense the other segments. Next, we introduce a feedback scheme and a selective diffusion strategy to keep the integrity of foreground objects, followed by a sticky trajectory method to remove noisy fragments and reduce blinking effect. Furthermore, an adaptive truncation strategy is introduced to raise the condensation ratio and improve the visual quality. Experimental results demonstrate the effectiveness of our approach.