A performance evaluation of keypoints detection methods SIFT and AKAZE for 3D reconstruction

In this paper, we discuss 3-dimensional reconstruction from multi-view images. In particular, we focus on SIFT[1] and AKAZE[2], well-known as keypoint detection/feature descriptor used for 3D reconstruction, and evaluate the performance of them. 3D reconstruction quality highly depends on the results of keypoint matching between images, so that it is important to evaluate the performance. First, this paper shows the results of keypoint matching by SIFT and AKAZE. Next, 3D reconstruction is performed based on each of the matching results, and the accuracy of 3D points is evaluated. We also use Multi View Stereo (CMVS[4]) to obtain dense 3D points. Moreover, in this study, we propose a new method for which SIFT and AKAZE are combined to obtain further detailed reconstruction, and show the better reconstruction results than the ones obtained by using only SIFT or AKAZE.