segmentation maps are bad (disconnected)

[mustachemo]

Describe the issue

I'm using blenderproc to make a dataset for a semantic segmentation task. I have to render the objects at different distances, so most of the time they are very small. The segmentation maps come out disconnected and it is not good for model training.

Minimal code example

def render_scene(scene_dir, distance="random", num_renders=100, resolution=3072):
    metadata = {}
    print(f"screne_dir: {scene_dir}")
    scene = glob(str(scene_dir) + "/*.blend")[0]

    # load the objects into the scene
    # blend_file = scene
    # objs = bproc.loader.load_blend(blend_file)

    # define a light and set its location and energy level
    light = bproc.types.Light()
    light.set_type("SUN")
    light.set_location(sample_sun_loc())
    field_of_view = np.deg2rad(0.4)
    # decrease the energy level to mimic d1-images
    light.set_energy(5)

    # define the camera intrinsics
    bproc.camera.set_intrinsics_from_blender_params(lens=field_of_view, lens_unit="FOV", clip_end=np.inf)
    bproc.camera.set_resolution(resolution, resolution)

    # Create camera poses for n renders. each camera pose corresponds to one image
    for i in range(num_renders):
        if type(distance) == str:
            distance = random.randint(50000, 200000)
            # distance = random.randint(10000, 20000)
            # distance = 5000
        position = np.random.uniform([-distance, -distance, -distance], [distance, distance, distance])
        x = position[0]
        y = position[1]
        z = np.sqrt(np.abs(distance**2 - x**2 - y**2))
        position = [x, y, z]
        quat, rot_matrix = create_tracking(position, [0, 0, 0])
        # print('position ', position, 'rot ', quat, rot_matrix.shape)
        matrix_world = bproc.math.build_transformation_mat(position, rot_matrix)
        bproc.camera.add_camera_pose(matrix_world)

    # activate depth rendering, render the whole pipeline
    # bproc.renderer.set_world_background([0,0,0])
    bproc.renderer.enable_segmentation_output(
        map_by=["category_id", "instance", "name"], default_values={"category_id": 1}, pass_alpha_threshold=0.01
    )

    data = bproc.renderer.render()
    satellite_name = scene.split("/")[-1].split(".")[0]

    try:
        os.mkdir("output/" + satellite_name)
    except OSError:
        print("Folder exists...")

    bproc.writer.write_hdf5("output/" + satellite_name, data)
    bproc.clean_up()

    for hdf5_file in glob("output/" + satellite_name + "/*.hdf5"):
        write_image(hdf5_file, "output")

Files required to run the code

No response

Expected behavior

BlenderProc version

4.2

[cornerfarmer]

Hey @mustachemo,

these seem to be aliasing effects of blenders raytracing. I am not sure there is anything we can do here.
One alternative would be to render the mask in a higher resolution and scale it down afterwards.

[mustachemo]

Hey @mustachemo,

these seem to be aliasing effects of blenders raytracing. I am not sure there is anything we can do here. One alternative would be to render the mask in a higher resolution and scale it down afterwards.

I see. Can I increase the number of rays to remedy that you think?
Also, your alternative solution just might work! Will try it and report back. Thanks for the reply