Hacker News Reader

Comment by aerodog | Hacker News Reader

is this substantially different from how we do 3d reconstruction from motion indoors?

It's approximately the same technique.

Outdoor 3D reconstruction is actually easier, because outdoors you have highly varied and detailed scenes, which enables dense feature matching and accurate dense depth map calculations. Also, you probably have decent GPS priors* and a preplanned flight path, which makes localization of each RGB image pretty easy.

* This is a big advantage that usually can't be replicated indoors. GPS is more accurate outside, and the significantly larger flight path means that GPS' main issue, low precision, isn't as much of a barrier.

GPS is more accurate because the location of the satellites is known to a high degree. If you knew the location of the WiFi access points to the same degree you could get pretty similar accuracy indoors. Apple indoor mapping gets 5-10m with dedicated staff walking the space & inputting their position on a mobile map (wide open spaces is where the biggest challenge is). If more APs adopt WiFi RTT then that's actually going to give you much more GPS-like behavior indoors.

I'm not sure how useful a 5-10 metre accuracy is going to be for indoor photogrammetry. 10 metres cubed is probably larger than the majority of indoor photo sets.

Now, for a military drone, I think they get better accuracy than 5 metres, but also the distances involved in an outdoor dataset mean that inaccuracies in the initial position estimate are much less significant.

You can get accuracy down to a few centimetres thanks to RTK Gnss. Most professional drones offer that now, though it is usually rather expensive.

It's also a question of the precision of timing. The GPS signal is entirely designed around precise timing, whereas Wifi isn't. There are a number of variables like buffer lengths that can be measured around to some extent, but are never going to be as precise.

Just using the very coarse RSSI signal Apple was getting 5-20m level accuracy depending on the venue 5 years ago & Google gets ~15m accuracy indoors worldwide (different mapping techniques).

You're thinking at the software level though. Look up 802.11mc though - that enables cm-level mapping.

Agreed, the GPS priors are very helpful -- the ability to specify prior known positions/orientations of photos, to some known error tolerances, is one of the power features in Agisoft Metashape Pro, which I wish was available in more photogrammetry software. Being able to know the actual nearby photos a priori and avoid the costly O(N^2) pairwise comparison is a great benefit.

Some research I did a few months ago was about trying to find ways to effectively bring in that advantage of outdoor photogrammetry into indoor environments, by piggybacking onto AR HMDs with built-in SLAM: https://www.youtube.com/watch?v=ldrGpGrOaZc

Reply on news.ycombinator.com