Creating a mesh for point cloud based on XYZ?

[slamfunk]

Hi guys,

so i'm looking at ways of using Kinect point cloud information to create 3D scans of things in Houdini. So far I am outputting XYZ .chan files from Processing (which communicates with the Kinect) and importing them in to Houdini via CHOPS (advice I saw in an earlier post on odforce). After doing this I find a nice set of points in my viewport which represent the scanned object. From here on how would I go about skinning the points?

I have tried to use a fluid surface object but the effect is too blobby and well.. fluidy.

Using a Skin SOP doesn't work cause there are no rows and columns amongst the points. Also there is no normal information.

Any pointers would be greatly appreciated.

Thanks.

P.s. Attached a picture of an example point cloud.

Edited March 22, 2012 by slamfunk

[anim]

you can use good old trick with flattening the pointcloud to camera plane then use triangulate2D sop to create mesh from them then move points back to original position

here is one of the threads where you can find more detailed description of the process

however flattening in Z may or may not for you, so make sure you chose correct axis or some arbitrary axis, best for you may be camera Z if your camera is matched witch your kinekt camera

[vi_rus]

Take a look at the PointCloudIso node.

[slamfunk]

Hey Tomas thanks for the heads up i'll give it a shot and see what I can create.

Sergei, I did take a look at the pointcloudiso node but I'm missing the point normal attribute. Is there a workaround for that you might know?

Edited March 22, 2012 by slamfunk

[edward]

Doesn't Kinect didn't estimate normals as well?

[petz]

Hey Tomas thanks for the heads up i'll give it a shot and see what I can create.

Sergei, I did take a look at the pointcloudiso node but I'm missing the point normal attribute. Is there a workaround for that you might know?

[ryew]

I don't know the answer, but the idea is really intriguing - please continue sharing your progress on odforce if you could

[zarti]

hi slamfunk ,

your picture shows points from a front projected scan ( only depth ) . if thats the case , shouldnt be hard to solve , imho .

in my attached file i tried to rebuild a full 3D pointCloud .

i couldnt find a way to totally fuse and clean the mesh , but at least transfered its normals to the pointCloud ..

--

the null1 SOP has few controls . hope this helps ..

-

-

remeshedzcan1.hipnc

[slamfunk]

Cool thanks for the advice guys, i'll have a look at your suggestions for the normals.

As per yesterdays tips I tried the triangulate2D sop but all that happened was Houdini crashed. I mean python started calculating once I activated t2d sop but after half an hour I escaped out and then Houdini crashed. Is it supposed to take that long? Should I just wait a little longer? I mean if it takes that long to cook one frame I must be doing something wrong right?

Also, i'm going on a little tangent here but is there anyway to ignore points that lie at 0,0,0(xyz) in the point cloud? I don't want to delete the points as the points are always changing but I just want to ignore the points that lie on the origin at every frame. Is there some sort of bounding box or threshold I can create for points?

Cheers

[anim]

you can create group from points you want, then use just that for calculation of your mesh

to ignore points at 0,0,0 you can create point group by expression like

length($TX, $TY, $TZ) > 0

or you can use bounding box or object to create group (Bounding Tab)

but in my opinion you can delete points you don't want after they came from CHOPs, it will not affect pointcloud calculation for other frames, because input for CHOPs will still be the same

you can do it similarly, in delete SOP by expression

length($TX, $TY, $TZ) < 0.0001

triangulate2D should be quite fast, depends on number of points you have

for example

10000 points took 0.6s to triangulate here

100000 points took 20s

1000000 points took 12m

since your points are just from one camera view, that approach should be fine and you'll get normals as well

[majikal]

hi..

i used brekel capture for capturing. You can get rgb, depth, pointcloud and obj surface out of it...

or you can just use raw depth map output from kinect as displacement texture on a grid and displace in direction of an camere (its slow and you need fov of kinect)...

point that i dont need like bgk i deleted manualy with conditions.

Edited March 23, 2012 by majikal

[slamfunk]

you can create group from points you want, then use just that for calculation of your mesh

to ignore points at 0,0,0 you can create point group by expression like

length($TX, $TY, $TZ) > 0

or you can use bounding box or object to create group (Bounding Tab)

but in my opinion you can delete points you don't want after they came from CHOPs, it will not affect pointcloud calculation for other frames, because input for CHOPs will still be the same

you can do it similarly, in delete SOP by expression

length($TX, $TY, $TZ) < 0.0001

triangulate2D should be quite fast, depends on number of points you have

for example

10000 points took 0.6s to triangulate here

100000 points took 20s

1000000 points took 12m

since your points are just from one camera view, that approach should be fine and you'll get normals as well

Thanks Tomas, that delete SOP with the expression did the trick! Got triangulate2D to work as well but the mesh is pretty nasty so I need to figure out a way to clean it up a little.

[slamfunk]

i used brekel capture for capturing. You can get rgb, depth, pointcloud and obj surface out of it...

Hey majikal thanks for the pointer. I had seen brekel kinect floating around but my objective was to get the information through Processing so I never followed it up. Nonetheless I had a look now and it's very handy!

[premini]

Hi, i revive this thread to ask if you finally achieved the results you were looking.... i am about to start the same path of action.

Thanks in advance!