SLAM in the wild
retained text
Being a hobby cyclist, I've wondered how well neural SLAM/SfM methods do on in-the-wild footage. So I mounted a camera on the bike and compare their track against GPS. Not that great. Most methods fail repeatedly, even when I reset them to the GT pose every 60 s so that they have a chance to recover in different conditions. LingBot-Map consistently outperforms the others across scenes. For SLAM people: do you think this is just distribution shift, or something about the LingBot-Maps' problem modeling? Any tips for a monocular SLAM worth trying that could work on such sequences? Or is it a case for Lidar or inertial SLAM systems? Intravilan scenes give better results, but many methods still break. On the contrary, the methods work for similar intravilan scene in LaMAR quite well.
document fetched 2026-10-31 00:23 · canonical https://reddit.com/r/computervision/comments/1xjpksm/slam_in_the_wild