The buildings are meticulously 3-D modeled and texture mapped--you can see the texture repetition artifacts. I've read some research about automatic modeling techniques from street view images+, but I'd suspect this was done by hand.
This is essentially the same as a traditional animated movie: gigantic sum of work to do, essentially infinitely parallel, and capable of being process-optimized such that two adjacent frames require much less work than freehanding them. They're probably using a large library of templated buildings (how many silhouettes are there at that resolution, really?), brushes, and then just applying detail work to the significant ones.
It's a big job, and it's impressive, but it isn't "We had enough people to freehand a scale map of the entire country, bwahaha."
The real question is, is this more useful than the photographic approach? i.e. does a user identify these landmarks better in this slightly abstracted hand drawn rendering than in a photo rendering of the same?
Is it fully 3D rendered? or just an isometric view like the old simcity? You can't rotate the viewpoint, as far as I've seen (but as I don't know Chinese I might be wrong).
Seriously by hand? How many mechanical turks?
Why do you say so? The textures and all seem to suggest otherwise. I am very much interested in the actual technology. Perhaps the textures were superimposed on the 3d models and not a direct input of the 3d modelling.
The research done by your fellow in MIT while in HKUST is also very impressive. Just that the result is not the same kind found in here - the one here is simply artistic model, not a realistic one.
The buildings are meticulously 3-D modeled and texture mapped--you can see the texture repetition artifacts. I've read some research about automatic modeling techniques from street view images+, but I'd suspect this was done by hand.
+ http://web.mit.edu/jxiao/Public/publication/2009/TOG/paper_h...
Looks like China's cheap labor is bridging the gap to AI.
This is essentially the same as a traditional animated movie: gigantic sum of work to do, essentially infinitely parallel, and capable of being process-optimized such that two adjacent frames require much less work than freehanding them. They're probably using a large library of templated buildings (how many silhouettes are there at that resolution, really?), brushes, and then just applying detail work to the significant ones.
It's a big job, and it's impressive, but it isn't "We had enough people to freehand a scale map of the entire country, bwahaha."
Perhaps they'll perform Searle's "Chinese room" experiment IRL.
5 replies →
The real question is, is this more useful than the photographic approach? i.e. does a user identify these landmarks better in this slightly abstracted hand drawn rendering than in a photo rendering of the same?
1 reply →
Is it fully 3D rendered? or just an isometric view like the old simcity? You can't rotate the viewpoint, as far as I've seen (but as I don't know Chinese I might be wrong).
Seriously by hand? How many mechanical turks? Why do you say so? The textures and all seem to suggest otherwise. I am very much interested in the actual technology. Perhaps the textures were superimposed on the 3d models and not a direct input of the 3d modelling.
The cost of labor in China is still low enough for this to happen.
The research done by your fellow in MIT while in HKUST is also very impressive. Just that the result is not the same kind found in here - the one here is simply artistic model, not a realistic one.