A Cross-Platform HD Dataset and a Two-Step Framework for Robust Aerial Image Matching

Image matching has been an active research area in the computer vision community over the past decades. Significant advances in image matching algorithms have attracted attention from many emerging applications. However, aerial image matching remains demanding due to the variety of airborne platforms and onboard electro-optic sensors, long operational ranges, limited datasets and resources, and constrained operating environments. We present two contributions in this work to overcome these challenges: a) an upgraded cross-platform image dataset built over images taken from an aircraft and satellite and b) a two-step cross-platform image matching framework. Our dataset considers several practical scenarios in cross-platform matching and semantic segmentation. The first step in our two-step matching framework performs coarse-matching using a lightweight convolutional neural network (CNN) with help from aircraft instantaneous parameters. In the second step, we fine-tune standard off-the-shelf image matching algorithms by exploiting
spectral
,
temporal
and
flow
features followed by cluster analysis. We validate our proposed matching framework over our dataset, two publicly available aerial cross-platform datasets, and a derived dataset using various standard evaluation methodologies. Specifically, we show that both steps in our proposed two-step framework help to improve the matching performance in the cross-platform image matching scenario.