Yeah and for niantic to achieve good photogrammetry with their random collection of photos taken from different angles, on different days, etc they would need some kind of ground truth to train on, which is implausible. You'd need to collect a parallel dataset of high-quality videos for traditional photogrammetry and .. hang on.
I was able to create a full 3d model of my window plant almost free of obscured areas from a few dozens still photos taken all around it, back in 2018, using the Capturing Reality photogrammetry app on a mobile i7-3610QM CPU with 8Gb RAM, in about 40-60 minutes.
And that's pretty mundane general public software, do we know for sure which algorithms are used by Niantic?
> and for niantic to achieve good photogrammetry with their random collection of photos taken from different angles, on different days, etc they would need some kind of ground truth to train on, which is implausible.
I'd say... the versatility of photos provides the "ground truth" on its own when combined to one single dataset. Say you want to program a guided drone shooting through urban areas, you want it to work under all sorts of conditions - day, night, rain, snow, the sun visible from all possible angles and throwing shadows.
A dataset that you can get from something like Street View? You can at best generate that once a year at enormous expense. Still valuable because a Street View car likely has a multitude of highest-quality GNSS receivers and possibly RTK navigation aids, but to make the dataset usable for 24/7/365 navigation you absolutely need a huge, huge amount of backfill.
Yeah and for niantic to achieve good photogrammetry with their random collection of photos taken from different angles, on different days, etc they would need some kind of ground truth to train on, which is implausible. You'd need to collect a parallel dataset of high-quality videos for traditional photogrammetry and .. hang on.
I don't understand why you insist on videos.
I was able to create a full 3d model of my window plant almost free of obscured areas from a few dozens still photos taken all around it, back in 2018, using the Capturing Reality photogrammetry app on a mobile i7-3610QM CPU with 8Gb RAM, in about 40-60 minutes.
And that's pretty mundane general public software, do we know for sure which algorithms are used by Niantic?
> and for niantic to achieve good photogrammetry with their random collection of photos taken from different angles, on different days, etc they would need some kind of ground truth to train on, which is implausible.
I'd say... the versatility of photos provides the "ground truth" on its own when combined to one single dataset. Say you want to program a guided drone shooting through urban areas, you want it to work under all sorts of conditions - day, night, rain, snow, the sun visible from all possible angles and throwing shadows.
A dataset that you can get from something like Street View? You can at best generate that once a year at enormous expense. Still valuable because a Street View car likely has a multitude of highest-quality GNSS receivers and possibly RTK navigation aids, but to make the dataset usable for 24/7/365 navigation you absolutely need a huge, huge amount of backfill.