Depth Adaptive Video Stitching

This paper proposes a novel approach to stitch videos fast and with high quality. In general, scene depth is frequently varying for dynamic video content. For example, moving foreground objects will move closely or far from a camera. The scene depth will be changing according to changed scene content. Therefore, accurate projection transform estimation corresponding to current scene depth is required in video stitching algorithms. However, projection transform estimation is a time-consuming task. To tackle this problem, we propose an algorithm to estimate projection transforms fast. The algorithm calculates initial projection transforms in a camera position calibration stage first. The projection transforms for successive frames are then obtained approximately based on a depth compensation method. In addition, a pixel indexing scheme is designed to speed up pixel projection between source video frames and panorama video frames. Using above optimization strategies, our video stitching algorithm can work fast and with high visual quality. Several real video stitching experiments are carried out. The experimental results show that our algorithm can improve stitching visual quality obviously with acceptable computing cost.