Visualizing The Matterhorn as an 860,000 point cloud | Hacker News Reader

Visualizing The Matterhorn as an 860,000 point cloud

mapbox.com

79 pointsbsudekum34 comments

I applaud the makers of this demo for adding a gif-based fallback on the error page for those of us who have no WebGL support. It shows:

"Your browser does not support this demo. Here is what you are missing:"

[https://www.mapbox.com/pointcloud/matterhorn/css/error.gif]

It is actually a raw data from UAV mapping; the whole story: https://www.mapbox.com/blog/sensefly-drone-mapping-matterhor...

I don't think so, the raw data should be much more accurate down to deci-, if no centimeters.

Height data derived from photogrammetry usually has lower vertical accuracy than that (even on fixed wing planes with better sensors).

How exactly do you justify disagreeing the blog post from the creators of the project? Do you believe they are lying about the data source?

The company behind the technology to turn images into point clouds is Pix4D [1]. Their website has another few pretty demos.

[1] pix4D.com

Doesn't work in Chrome/Chrome Canary/Safari on OS X. Just get a black screen after the points are loaded

Works on Chrome on 10.8 here. But not Safari.

I'm still on 10.7, Chrome 32.0.1664.3 dev

Did not work on Chrome 30/Mac OS X 10.8.5 for me.

I got just the black screen on my office iMac with an AMD card, but it works fine on my much older Macbook with an nvidia card.

Took a reload to get it chugging on my firefox. Win7 64.

it's just as fun to flip it around and explore the big hole in the ground

This crashes Safari on my mac. Seems to be a segfault originating somewhere in the graphics card driver. Card is an ATI HD 4870. Running MacOS X 10.8.5/Safari 6.0.5.

Also crashes in Firefox 24 on the same machine.

worked for me in Safari on my Retina MacBook Pro. Running OS X 10.8.5 and Safari 6.0.5

Yes, I think it must be a problem with the graphics card or driver rather than the browser itself.

Also Crashes Ff on my mbp early 2011 with 10.7.5

This is a low res version of the original data, because of limitations of webgl. If you want to see the making of, as well as the full res pointcloud (over 300 millions!!!), check out this video:

Mapping the impossible: http://youtu.be/jh4kRatBNkk

Wrong link, the video mapping the impossible is this one: http://youtu.be/NuZUSe87miY

Can we get Leap Motion support in this? Their API is pretty simple/well made, and it's very well suited to easily rotating 3D stuff.

i don't have a reference to compare it to aside from a heightmap i used for some rendering years ago, but 25mb for the *.asc seems pretty big for the quality it puts out

What's the point of point clouds?

my question also. This is useful for what, exactly?

Point clouds are not very useful themselves, if we had a choice we would skip them. They are the result of turning any measured data (like pictures or depth samples or whatever) into a 3d visualisation.

Ideally you have collected enough data to reconstruct the surface geometry with reasonable accuracy, but realistically often you don't.

Our minds are rather good at inferring surface geometry from point clouds, our algorithms not so much yet as far as I know. So sometimes (as in the case of this visualisation) it suffices to show people just the point cloud.

Because it's there.

Point cloud data is (relatively) cheap and easy to collect. However turning point clouds into 3D surfaces is computationally Very Hard and even the very best tools still require a lot human intervention to give acceptable results. Having good tools that are able to work with point clouds directly dramatically speeds up the time from data collection to data visualization and being able to actually use the data.

Hmm, why is it very hard? Wouldn't just connecting the points between them yield acceptable results? I can understand if the points are disparate, but this dataset is pretty much a grid, and it looks like you could easily generate the surface from the points.

What am I missing?

How do you decide which points to connect?

Use an unstructured grid (triangulation) to connect the elements. This is routinely used in numerical modelling (think computational fluid dynamics) to generate linked data sets from scatter data (such as this) in 2 and 3D (i.e. a surface and a volume).

That will work in about 50-70% of cases in the real world depending on the quality of your data and how sensitive you are to errors. For the remaining 30-50% you have to fix things by hand. The further problem is there is no good heuristics to separate the cases where it worked from where it didn't, so you still have to go through all the data to flag the cases which have to be fixed. Trust me, if you can come up with a way to make it work in general, you could build a very successful company around selling that software.

What phireal said. If it's a sparse point cloud, sure, it would be troublesome. The one in the example is pretty dense, and it's almost a grid, so just connect neighboring points together, no?

How do you decide which points are neighboring?

Its actually fairly interesting technology. A while ago (years?) there was a post about a graphics engine which provided unlimited detail. You could literally zoom from 50,000 feet down to observe a single blade of grass. That technology is still to do anything but the demo's are impressive.

https://www.youtube.com/user/EuclideonOfficial

Reply on news.ycombinator.com