Image Matching in Deep Learning Era:Methods,Applications and Challenges
Image matching is a crucial technique within the field of computer vision,primarily focused on identifying and establishing point correspondences between two different images depicting the same scene.It seeks to find points in one image that correspond to points in another,thus enabling a wide range of computer vision tasks that rely on the analysis of multiple images of the same object or scene from different viewpoints or at different times,including but not limited to 3D reconstruction,motion tracking,image stitching for panoramic views,and visual localization.Traditionally,this process has leaned heavily on the use of hand-crafted keypoint detectors and local descriptors,i.e.,algorithms and methodologies designed to pinpoint and describe discriminative features within a local image region,aiming to achieve invariance to scale,rotation,and changes in lighting and perspective.In recent years,with the revolutionary development of deep learning in many areas of computer vision,image matching methods have switched from handcrafted design style to relying on deep learning.The advent of deep learning technologies has catalyzed significant advancements in the area of image matching,and numerous deep learning based image matching techniques have emerged,showcasing promising results across a wide range of benchmarks.This has also significantly accelerated the development of many downstream applications of image matching,notably including structure from motion,visual localization,and simultaneous localization and mapping(SLAM),among others.This paper aims to provide a comprehensive overview of deep learning-based image matching methods that have emerged in recent years.By delving into the core challenges of image matching,including keypoint detection,local feature description,dense matching,and mismatch removal,it offers a detailed summary of the innovative deep learning approaches devised to tackle these issues.This systematic review not only highlights the advancements in the field but also sheds light on how these cutting-edge methods have redefined the landscape of image matching,setting new benchmarks for accuracy,efficiency,and reliability.Specifically,it first delineates the problem definition of image matching and describes the main challenges.Then,it proceeds to dissect each problem associated with image matching,offering a thorough analysis of typical and emblematic methods.Additionally,it delves into the critical techniques employed by deep learning to address these issues,providing an in-depth exploration of how these innovative approaches can effectively solve the challenges inherent in image matching.Moreover,some highly related downstream tasks of image matching are described along with a detailed introduction of their state of the art.These downstream tasks include 3D reconstruction/structure from motion,image based localization,and simultaneous localization and mapping.Besides exploring these downstream applications,this paper provides a comprehensive description of popular benchmarks for image matching and its downstream tasks.Finally,the paper discusses the remaining challenges and future research directions.In conclusion,this paper presents itself as an invaluable resource for researchers and engineering technicians within related fields,enabling the swift assimilation of knowledge concerning the fundamentals,challenges,key technological advancements,and the current state of the art in image matching.As such,it can be served as a comprehensive resource for researchers venturing into this field,providing references in terms of research directions and dataset resources.Through its detailed exposition,the paper aims to catalyze further exploration and innovation,thereby could contributing significantly to the advancement of image matching and its application in advancing the frontiers of computer vision.
image matchfeature point matchdense match3D reconstructionvisual localizationsimultaneous localization and mappingdeep learning
NETL
NSTL
万方数据
