How useful is spatial data over time, does it decay or age much?
Is the geographical data more useful, or are buildings and other structures more important?
Genuinely don't know much in this space.
How useful is spatial data over time, does it decay or age much?
Is the geographical data more useful, or are buildings and other structures more important?
Genuinely don't know much in this space.
Its using Visual descriptors to generate a pointcloud. Buildings and text are really great for creating descriptors, so when they change you loose key points for "localizing"(ie getting your position). This needs to be updated as those buildings change.
You also need a day/night dataset (although some newer descriptors are day/night resistant)
It's easier to take a look via change of dates in google street view, they have almost 20 years coverage. You can see how the data ages and decays or doesn't because it's tied to the place it represents.
Shops come and go, churches do not move, schools tend not to move much, industry areas is somewhat dynamic, military installations might be static or dynamic, trees grow or are removed.
Yes. Even dealing with "what does it look like in rain? What does it look like in snow?" is hard. Hell... "What does it look like at night" is hard. Hell.... what does it look like at noon vs sundown (no shadow vs long shadows) is hard.
Have you ever seen a commercial use of anything like this? That should give you a hint about how reliable these systems get.
It's the combination of geographical data (maps) linked to its visual representation in the world (footage of structures, roads, landscape features) that is useful.
The geographical data already exists in digital maps. And I would expect competent militaries already have maps of enemy territory. It's the second part that was so far missing.
This combined set allows the training of AI models that can say, "When my surroundings look like x, that looks like y on a map".
So when your drone's GPS gets jammed, it can look at its surroundings, reference its (internal and offline) maps, figure out where it is, and navigate.
Compred to what? Datasets at this scale are rare. You're not comparing against another ideal dataset, you're comparing against having nothing.
There are so many companies these days doing recording for self driving cars and/or street view like applications. Also sites like Flickr collect huge sets of geo tagged photos, as do companies like Meta where tons of geo tagged images are shared each day via their different outlets.
Niantic has the benefit that they can steer "volunteers" to specific points, though.
Maybe Niantic intentionally steered players towards remote locations only reachable on foot, understanding that this data is more scarce and therefore more valuable?
I don't what class of models they use here, specifically, but a generic classifier shouldn't depend on a single feature. And neighbourhoods don't typically get razed or remodeled/painted over in a fortnight.
... Except, well, when it's the doing of this same, so called "defence" industry.