Q&A Explanation

[themasterlink]

Hey,

we are happy that you use BlenderProc and that you stumbled upon this Q&A section. You can always ask a question here and see if someone can help you.

But, before you ask your question make sure that you checked the documentation and all answered questions if your question was already answered before.

Best,
Max

[daeyeoplee]

[themasterlink]

Could you please open a new discussion thread?

Then I will tell you the answer ;)

[deephog]

I'm really new to BlenderProc, and I found it to be an amazing tool. One thing I'm curious is that, can I add Human objects into the scene and can I manipulate the Human object in a natural way? Thanks!

[themasterlink]

Hey there,

have you checked the Amass example?

That is the closest thing we have to human rendering.

Best,
Max

[deephog]

Hi, thank you for your advice! I went through AMASS example, downloaded and organized the dataset as instructed, but I couldn't find where the "taxonomy.json" file comes from, and creating a new empty file with the name apparently doesn't work. Could you provide me somemore info of how to find it? Thanks!

[themasterlink]

Hey,

you can find that inside the resources folder:

https://github.com/DLR-RM/BlenderProc/blob/main/blenderproc/resources/AMASS/taxonomy.json

Best,
Max

[Sodek93]

Hey, is it possible to reuse convex_decompositions objects in the scene instead of reloading it each time it is used ? convex decomposition is done only once - thats fine!

For example i have 10 times the same object with different position and orientation in my scene. Currently loading the convex_decomposition from folder is done for every 10 objects, which takes some time.

For loading the objects themselfs i use the cached_objects variable (ObjectLoader.load_obj()) which works fine.

When calling build_convex_decomposition_collision_shape(vhac_pathvariable) all i pass to function is the vhac_path. The cache_dir has the default value.

Is there a way to write the convex_decomposition into the cached_objects dict or any other way to prevent reloading the same convex_decompositions into the scene?

Best regards,
Sodek

[Sodek93]

Hey,
i'll update blenderproc and see if it works. I am using version 2.3.0 right now.

I'll try your suggestion with first doing the convex decomposition and then copy the object. I'll give you feedback afterwards.

Thanks for your support!

[themasterlink]

Hey,

I tried to add the cache_objects property to the decompose function. If you now use it like this:

build_convex_decomposition_collision_shape(args.vhacd_path, cached_objects=cached_objects)

I had no time to test it thoroughly, but you can do that ;)

https://github.com/DLR-RM/BlenderProc/tree/test_decompose_with_caching

Best,
Max

[Sodek93]

I'll test and give you feedback soon :-)

[Sodek93]

Hey Max,
your solution removes the reloading of the convex decomposition - but i found out that my (time) problem is not the loading but the handling of the convex parts itself... sorry for that!

In MeshObjectUtility.build_convex_decomposition_shape()
line 239 to line 24x(i got blank lines beteween) which takes convex parts, parents it with the object itself, enable_rigidbody and hide the convex parts takes aprox. 20 sec for 40 convex decomposition parts of one object.

Is this a local problem for just me?

I never thought that this could use so much ressources.

Here the code that takes about 20 sec for a 5000 vertices Object with 40 convexed parts:

    for part in parts:
        part.set_parent(self)
        part.enable_rigidbody(True, "CONVEX_HULL")
        part.hide()

Using new version of blenderproc and using cached_objects takes script runtime from 1200 sec to 950 sec. (i use about 50 times the same object)

I'll test the function 3 times with one of the lines commented out.

I think the main part in reducing the runtime was the new .obj loader?

[themasterlink]

Hey,

this gets too complicated for a discussion, could you open an issue.

Best,
Max

[kgrouleff]

Hey Max,

I'm trying to simulate various camera poses and light settings using blenderproc. So far I can change the camera poses, but when I try to change the lights all scenes seems to get the last light value that is geenrated in the loop. My code looks something like this:

light = createLight(sourceType=LightSource.SUN)

def changeCameraPose(camera, cameraPose):
    print("Changing camera pose:")
    rotation = [cameraPose['rot'][0], cameraPose['rot'][1], cameraPose['rot'][2](*rotationValues[camera])]
    matrix_world = bproc.math.build_transformation_mat(cameraPose['pos'], rotation)
    bproc.camera.add_camera_pose(matrix_world)

def changeLightSettings(light):
    # Setting random light poses
    light.set_rotation_euler([round(random.uniform(4.3, 5.8), 1),
                              round(random.uniform(-1.3, 2.8), 1),
                              round(random.uniform(-0.5, 0.5), 1)],
                             )

    energy = random.uniform(-20, 200)
    print(f"Light energy: {energy}")
    light.set_energy(energy)

def main():
   
   for j in range(settings.numOfRenders):

        # Changing light and camera poses
        for camera, cameraPose in enumerate(cameraPoses.values()):
            print(f"Generating images for camera: {camera + 1}/{len(cameraPoses)}")
            # Modify camera and light - cameraPos is a dict, with keys: pos and rot
            changeCameraPose(camera, cameraPose)
            changeLightSettings(light)

data = bproc.renderer.render()

How should I apply different light to the sceens? I have already looked at the examples, but I can't find pointers on how to do this?

[themasterlink]

Hey,

you have to use key frames, see the example for keys frames.

If you need more help, feel free to ask.

Best,
Max

[kgrouleff]

Hi Max, I will take a look at that. Thanks!

[themasterlink]

Here is the correct link:

https://dlr-rm.github.io/BlenderProc/docs/tutorials/key_frames.html#objects