Mining Cross-Location Image "Hard Doppelgangers" at Scale

gabriberton · x · 2026-07-13

Challenging the traditional view that similar image pairs (hard doppelgangers) in visual localization must come from the same building, a developer has provided counterexamples. By mining the SF-XL dataset, they discovered numerous image pairs separated by over 200 meters that share extremely similar visual features. The author noted that by using open-source street view imagery like Mapillary and combining it with MegaLoc for similarity retrieval, one can mine an almost infinite number of such cross-location image pairs, thereby providing rich training data for computer vision models.

Related event: Hard Doppelgangers Disrupt 3D Reconstruction(2 posts)→

Original post →

More from Research

Research channel →