Large-scale outdoor augmented reality scenes using camera pose based on learned descriptors

patent

US11568642B2

The United States Patent and Trademark Office, 600 Dulany Street, Alexandria, VA, US

required - licence fee

VUT, Adobe

Lukáč Michal (Adobe)
Wang Oliver (Adobe)
Brejcha Jan, Ing., Ph.D. (DCGM FIT BUT)
Hold-Geoffroy Yannick (Adobe)
Čadík Martin, doc. Ing., Ph.D. (DCGM FIT BUT)

camera pose estimation, outdoor scenes, augmented reality, learned descriptors

We introduce a solution to large scale Augmented Reality for outdoor scenes by registering camera images to textured Digital Elevation Models (DEMs). To accommodate the inherent differences in appearance between real images and DEMs, we train a cross-domain feature descriptor using Structure From Motion (SFM) guided reconstructions to acquire training data. Our method runs efficiently on a mobile device and outperforms existing learned and hand-designed feature descriptors for this task.