Reference Pose Generation for Long-term Visual Localization via Learned Features and View Synthesis.

Plotting the reprojection error (after nonlinear optimization) against the number of effective inliers, we observe a clear trend across different features: D2-Net recovers the most matches, followed by R2D2 and SIFT; while SIFT features were most accurately localized in the images, D2-Net has the largest reprojection errors.

also in 19,460 other papers