Spherical Feature Extraction:Benchmark And Evaluation

With the rapid development of various technologies in the field of computer vision,the research and use of panoramic images with a wide field of view has gradually increased.The panoramic image is captured by an omnidirectional camera(such as a reflective imaging system or a fisheye imaging system)or by stitching multiple fluoroscopic images.Its full coverage field provides a rich source of image features and increases the likelihood of matching features between views.At present,panoramic images have been successfully applied to many new applications.Feature detection and matching problems have been one of the basic problems in computer vision research,and research in panoramic images has gradually increased.The application has certain requirements on the performance of panoramic feature extraction.However,compared with the planar feature extraction,the evaluation of the panoramic feature algorithm has no unified evaluation benchmark,and fails to achieve the uniform measurement of the advantages and disadvantages of each panoramic feature algorithm.In order to solve this problem,this paper designs and implements the evaluation benchmark of the spherical panoramic image feature algorithm by studying and analyzing the evaluation criteria of the planar feature algorithm.The evaluation benchmark includes a virtual scene data set generated under influencing factors such as rotational translation and corresponding depth information,and measures the repetition rate and matching accuracy rate of the feature key points in the existing feature detection and matching algorithm based on the panoramic image.Perform an evaluation analysis.This paper generates a panoramic dataset and its corresponding depth information in a virtual scene.The dataset contains panoramic data under various influencing factors,and proposes a method to recover the metrics by using the depth information to recover the 3D point cloud to evaluate the metrics.Algorithm,and finally compare and analyze the two feature extraction algorithms according to the proposed evaluation benchmark.