Comparative Evaluation of Hand-Crafted and Learned Local Features

摘要: Matching local image descriptors is a key step in many computer vision applications. For more than decade, hand-crafted such as SIFT have been used for this task. Recently, multiple new learned from data proposed and shown to improve on terms of discriminative power. This paper dedicated an extensive experimental evaluation features establish single protocol that ensures comparable results. In matching performance, we evaluate the different regarding standard criteria. However, considering performance isolation only provides incomplete measure quality. example, finding additional correct matches between similar images does not necessarily lead better when trying match under extreme viewpoint or illumination changes. Besides pure descriptor matching, thus also context image-based reconstruction. enables us study set practical criteria including retrieval, ability register strong changes, accuracy completeness reconstructed cameras scenes. To facilitate future research, full pipeline made publicly available.