KermM: Note that I used the past tense. You said it is fixed now, I am curious what the fix was that resulted in that being the current state of affairs.

[edit]
oh ick, sounds like an unpleasant thing to track down.

This is awesome, this is what I dreamed of, real cities in the MC. I hope in success of this project. Please make it real, make it huge. Publish donation buttons or create a small kickstarter campain (you already have a prototype). I'll support you with joy! Cheers.

Thanks! It actually is completely functional, as you point out, but getting the data for the buildings to build the cities, plus writing more parallelization tools to make the building conversion as parallel as possible, is a real sticky wicket. Having gotten very little traction from the Google Geo team, I'm currently trying to reach out to Here.com and Bing Maps as possible sources for building models.

That's probably preferable to doing it manually from a bunch of images

Taking a vision class this semester and am working on the more algorithmically interesting (i.e. not just voxelizing) version of this again.

In the process of building a window/door/balcony detectors with Scala + OpenCV. The next step will be clustering the detected windows/doors/balconies to detect common submodels. Then "snapping to grid" the detected components within the image, and then probably applying a reinforcement-learning agent to optimize a tradeoff between realism/piece-count.

My SVM is misbehaving for detection, however given labels, the clustering+averaging works reasonably well. There's some small problems for now with the aspect-ratio binning because the hand-labeled stuff wasn't labeled carefully enough, but that should be solved when the detector is working and handing out its own labels of things with pre-binned ratios.



I also encountered some hilarity with pushing the limits of shadow-layering in OS X's window manager while debugging some stuff in an inner-loop by popping up windows with current image state:

Snapping found objects to a non-overlapping grid works now:


Here it is working on the test images from the previous post too:

Double post, but I don't care, because this is cool:



I appear to have an OBOE in one of my boundary conditions or something, because of those couple of random stripey bits, but otherwise this is exactly what the region decomposition should look like.

Updated status, if anyone is interested in technical stuff. Otherwise...

Whoa, awesome work! That looks like something you could easily build out of LEGO now. Based on your comments in our staff chat channel, though, I must ask you how expensive that computation was, in big-O terms and in realtime terms.

The brick layout is, at this point, the bottle neck. For smaller-scale images, it's pretty close to a tie between that and the constraint-program solver for the snap-to-grid feature since the constraints become harder and the dynamic program for the bricks becomes easier. Without doing any formal analysis, I think it's something like O(w^2 * h^2 * b * c) for an w * h image patch, with b bricks and color palette of size c. The part of the search that has a linear dependence on c could probably be replaced with something more complicated but more efficient, like a kNN regression. In terms of real-time performance, it was definitely a couple minutes, the bulk of which was the top third, which has no holes to subdivide the DP:


Each of those patches is handled separately.

*necrobump* I can't talk a whole lot about it, but I wanted to mention that I've been working a little on an aspect of this project again. Here's a screenshot (for example) of an early representation of the MetLife building, Grand Central terminal, and surrounding areas. As you can see, these are individual voxels, not solids.

Why does it have all the empty space?

The type of data capture that this data comes from only records points on solid surfaces (like trees, cars, buildings, the ground, water...); it's up to algorithms that I'm researching to make volumes out of all these points.

Hi all,
I've been admiring your project for a while, and I thought it was dormant. I tried your dynmap link, and it was online now and it looks awesome! I'm working on a similar project, although not nearly as large in scope. I'm generating a model of Northampton, Massachusetts where I live. I created a 1:1 scale mono-tone model from LIDAR (like your building above), but now we're going through and recoloring everything by hand (mostly) to modify the buildings to look accurate to real life, and correct imperfections in the data. Now that I see you are working on this project again, I was wondering if maybe you would be willing to talk to me about how your project actually works from a technical standpoint. I was talking to the head of the Google Computer Vision team yesterday, and he was intrigued in the idea of automating my project even further, and coloring the buildings from data.
Oh and by the way I'm a high school senior, and I completed the generation of my project by myself with pretty much no prior knowledge of lidar and GIS data. I do know a lot about Minecraft though.
Additionally, it would be awesome if I could be whitelisted on your server so I can look at it more closely, to see how I could improve my Minecraft city model.
You can check out my website (it hasn't been updated in a while though) Geominecraft.com. if you want, you can also check out the live map of my project at: http://162.244.165.71:1150.
my email is zev123@gmail.com, and I'd love to get in touch with you about your project and how I could automate mine.

Congratulations; I'm very impressed what you've done! I must admit that when I started this project, I had almost no idea what I was doing either; Jack Twiley, the creator of TopoMC, guided me most of the way to where I am today in terms of understanding geographic data, GDAL, and so on. I did just finish my PhD in distributed systems, though, so I have an edge in creating parallelized systems to manage and process data.

It was, it is, and thank you! Unfortunately, because of what I'm working on this for, I can't say a great deal about the internals of the system, but I'll provide any guidance that I can.

That bodes well for my project. Thank you for sharing that anecdote. As I said, I'll share what I can, but some of it is still in flux, and a lot of it is still under wraps.

It looks great! The manual adjustments to the buildings and the (manual?) generation of the trees looks particularly impressive.
Do feel free to email me; my email address is christopher at this website's domain name.

I saw your message on your server...
Do you think you could Skype sometime? We have our tech class from 10:30 to 12:30.
I also did see all of your responses, I was going to send you an email, but I haven't gotten around to doing so yet.

No problem, whenever you're ready. As I said, I'll do my best to tell you about how the system I'm building works without giving away the secret sauce. I'd be happy to Skype as well if that would help your project; 11:30am or later is best for me.

Found a few things out about the SparseWorld dynmap as I tried to access it and it was not the happiest experience. By default, the spawn is a negative x coordinate, instead of a positive y coordinate, and when manually entering the coordinates, the area is completely black.

It'd be cool to see NYC from overhead, just some bug fixes need to occur