Diverse scene stitching from a large-scale aerial video dataset

Diverse scene stitching is a challenging task in aerial video surveillance. This paper presents a hybrid stitching method based on the observation that aerial videos captured in real surveillance settings are neither totally ordered nor completely unordered. Often, human operators apply continuous monitoring of the drone to revisit the same area of interest. This monitoring mechanism yields to multiple short, successive video clips that overlap in either time or space. We exploit this property and treat the aerial image stitching problem as temporal sequential grouping and spatial cross-group retrieval. We develop an effective graph-based framework that can robustly conduct the grouping, retrieval and stitching tasks. To evaluate the proposed approach, we experiment on the large-scale VIRATaerial surveillance dataset, which is challenging for its heterogeneity in image quality and diversity of the scene. Quantitative and qualitative comparisons with state-of-the-art algorithms show the efficiency and robustness of our technique.