Evaluating hand-crafted and learning-based features for photogrammetric applications

The image orientation (or Structure from Motion – SfM) process needs well localized, repeatable and stable tie points in order to derive camera poses and a sparse 3D representation of the surveyed scene. The accurate identification of tie points in large image datasets is still an open research topic in the photogrammetric and computer vision communities. Tie points are established by firstly extracting keypoint using a hand-crafted feature detector and descriptor methods. In the last years new solutions, based on convolutional neural network (CNN) methods, were proposed to let a deep network discover which feature extraction process and representation are most suitable for the processed images. In this paper we aim to compare state-of-the-art hand-crafted and learning-based method for the establishment of tie points in various and different image datasets. The investigation highlights the actual challenges for feature matching and evaluates selected methods under different acquisition conditions (network configurations, image overlap, UAV vs terrestrial, strip vs convergent) and scene's characteristics. Remarks and lessons learned constrained to the used datasets and methods are provided.