Annotators Information

Annotators provide ground truth annotations corresponding to the rendered scene. Within replicator, there are multiple off-the-shelf annotators, each of which is described below.

Annotator Registry

The annotator registry is where all of the Annotators are. To get access to them, you can use the function omni.replicator.core.annotators.get().

The current annotators that are available through the registry are:

Annotators
Standard Annotators	RT Annotators	PathTracing Annotators
LdrColor/rgb	SmoothNormal	PtDirectIllumation
HdrColor	BumpNormal	PtGlobalIllumination
camera_params/CameraParams	Motion2d	PtReflections
normals	DiffuseAlbedo	PtRefractions
motion_vectors	SpecularAlbedo	PtSelfIllumination
cross_correspondence	Roughness	PtBackground
distance_to_image_plane	DirectDiffuse	PtWorldNormal
distance_to_camera	DirectSpecular	PtRefractionFilter
primPaths	Reflections	PtMultiMatte<0-7>
bounding_box_2d_tight_fast	IndirectDiffuse	PtWorldPos
bounding_box_2d_tight	DepthLinearized	PtZDepth
bounding_box_2d_loose_fast	EmissionAndForegroundMask	PtVolumes
bounding_box_2d_loose	AmbientOcclusion	PtDiffuseFilter
bounding_box_3d_360		PtReflectionFilter
bounding_box_3d_fast
bounding_box_3d
semantic_segmentation
instance_segmentation_fast
instance_segmentation
skeleton_data
pointcloud
CrossCorrespondence
MotionVectors

Some annotators support initialization parameters. For example, segmentation annotators can be parametrized with a colorize attribute specify the output format.

omni.replicator.core.annotators.get("semantic_segmentation", init_params={"colorize": True})

To see how annotators are used within a writer, we have prepared scripts that implement the basic writer which covers all standard annotators. How to get there is shown in Scripts for Replicator.

Standard Annotators

These annotators can be used in any rendering mode. Each annotator’s usage and outputs are described below.

LdrColor

Annotator Name: LdrColor, (alternative name: rgb)

The LdrColor or rgb annotator produces the low dynamic range output image as an array of type np.uint8 with shape (width, height, 4), where the four channels correspond to R,G,B,A.

Example

import omni.replicator.core as rep

async def test_ldr():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    cone = rep.create.cone()

    cam = rep.create.camera(position=(500,500,500), look_at=cone)
    rp = rep.create.render_product(cam, (1024, 512))

    ldr = rep.AnnotatorRegistry.get_annotator("LdrColor")
    ldr.attach(rp)

    await rep.orchestrator.step_async()
    data = ldr.get_data()
    print(data.shape, data.dtype)   # ((512, 1024, 4), uint8)

import asyncio
asyncio.ensure_future(test_ldr())

Normals

Annotator Name: normals

The normals annotator produces an array of type np.float32 with shape (height, width, 4). The first three channels correspond to (x, y, z). The fourth channel is unused.

Example

import omni.replicator.core as rep

async def test_normals():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    cone = rep.create.cone()

    cam = rep.create.camera(position=(500,500,500), look_at=cone)
    rp = rep.create.render_product(cam, (1024, 512))

    normals = rep.AnnotatorRegistry.get_annotator("normals")
    normals.attach(rp)

    await rep.orchestrator.step_async()
    data = normals.get_data()
    print(data.shape, data.dtype)   # ((512, 1024, 4), float32)

import asyncio
asyncio.ensure_future(test_normals())

Bounding Box 2D Loose

Annotator Name: bounding_box_2d_loose

Outputs a “loose” 2d bounding box of each entity with semantics in the camera’s field of view.

Loose bounding boxes bound the entire entity regardless of occlusions.

Output Format

The bounding box annotator returns a dictionary with the bounds and semantic id found under the data key, while other information is under the info key: idToLabels, bboxIds and primPaths.

{
    "data": np.dtype(
                [
                    ("semanticId", "<u4"),          # Semantic identifier which can be transformed into a readable label using the `idToLabels` mapping
                    ("x_min", "<i4"),               # Minimum bounding box pixel coordinate in x (width) axis in the range [0, width]
                    ("y_min", "<i4"),               # Minimum bounding box pixel coordinate in y (height) axis in the range [0, height]
                    ("x_max", "<i4"),               # Maximum bounding box pixel coordinate in x (width) axis in the range [0, width]
                    ("y_max", "<i4"),               # Maximum bounding box pixel coordinate in y (height) axis in the range [0, height]
                    ('occlusionRatio', '<f4')]),    # Occlusion (visible pixels / total pixels), where `0.0` is fully visible and `1.0` is fully occluded. See additional notes below.
                ],
    "info": {
        "idToLabels": {<semanticId>: <semantic_labels>},    # mapping from integer semantic ID to a comma delimited list of associated semantics
        "bboxIds": [<bbox_id_0>, ..., <bbox_id_n>],         # ID specific to bounding box annotators allowing easy mapping between different bounding box annotators.
        "primPaths": [<prim_path_0>, ... <prim_path_n>],    # prim path tied to each bounding box
    }
}

Note

bounding_box_2d_loose will produce the loose 2d bounding box of any prim in the viewport, no matter if is partially occluded or fully occluded.

occlusionRatio can only provide valid values for prims composed of a single mesh. Multi-mesh labelled prims will return a value of -1 indicating that no occlusion value is available.

Example

import omni.replicator.core as rep
import omni.syntheticdata as sd

async def test_bbox_2d_loose():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    cone = rep.create.cone(semantics=[("prim", "cone")], position=(100, 0, 0))
    sphere = rep.create.sphere(semantics=[("prim", "sphere")], position=(-100, 0, 0))
    invalid_type = rep.create.cube(semantics=[("shape", "boxy")], position=(0, 100, 0))

    # Setup semantic filter
    sd.SyntheticData.Get().set_instance_mapping_semantic_filter("prim:*")

    cam = rep.create.camera(position=(500,500,500), look_at=cone)
    rp = rep.create.render_product(cam, (1024, 512))

    bbox_2d_loose = rep.AnnotatorRegistry.get_annotator("bounding_box_2d_loose")
    bbox_2d_loose.attach(rp)

    await rep.orchestrator.step_async()
    data = bbox_2d_loose.get_data()
    print(data)
    # {
    #   'data': array([
    #       (0, 443, 198, 581, 357, 0.0),
    #       (1, 245,  92, 375, 220, 0.3823),
    #       dtype=[('semanticId', '<u4'),
    #              ('x_min', '<i4'),
    #              ('y_min', '<i4'),
    #              ('x_max', '<i4'),
    #              ('y_max', '<i4'),
    #              ('occlusionRatio', '<f4')]),
    #   'info': {
    #       'bboxIds': array([0, 1], dtype=uint32),
    #       'idToLabels': {'0': {'prim': 'cone'}, '1': {'prim': 'sphere'}},
    #       'primPaths': ['/Replicator/Cone_Xform', '/Replicator/Sphere_Xform']}
    #   }
    # }

import asyncio
asyncio.ensure_future(test_bbox_2d_loose())

Bounding Box 2D Tight

Annotator Name: bounding_box_2d_tight

Outputs a “tight” 2d bounding box of each entity with semantics in the camera’s viewport.

Tight bounding boxes bound only the visible pixels of entities. Completely occluded entities are ommited.

Output Format

The bounding box annotator returns a dictionary with the bounds and semantic id found under the data key, while other information is under the info key: idToLabels, bboxIds and primPaths.

{
    "data": np.dtype(
                [
                    ("semanticId", "<u4"),          # Semantic identifier which can be transformed into a readable label using the `idToLabels` mapping
                    ("x_min", "<i4"),               # Minimum bounding box pixel coordinate in x (width) axis in the range [0, width]
                    ("y_min", "<i4"),               # Minimum bounding box pixel coordinate in y (height) axis in the range [0, height]
                    ("x_max", "<i4"),               # Maximum bounding box pixel coordinate in x (width) axis in the range [0, width]
                    ("y_max", "<i4"),               # Maximum bounding box pixel coordinate in y (height) axis in the range [0, height]
                    ('occlusionRatio', '<f4')]),    # Occlusion (visible pixels / total pixels), where `0.0` is fully visible and `1.0` is fully occluded. See additional notes below.
                ],
    "info": {
        "idToLabels": {<semanticId>: <semantic_labels>},    # mapping from integer semantic ID to a comma delimited list of associated semantics
        "bboxIds": [<bbox_id_0>, ..., <bbox_id_n>],         # ID specific to bounding box annotators allowing easy mapping between different bounding box annotators.
        "primPaths": [<prim_path_0>, ... <prim_path_n>],    # prim path tied to each bounding box
    }
}

Note

bounding_box_2d_tight bounds only visible pixels.

occlusionRatio can only provide valid values for prims composed of a single mesh. Multi-mesh labelled prims will return a value of -1 indicating that no occlusion value is available.

Example

import omni.replicator.core as rep
import omni.syntheticdata as sd

async def test_bbox_2d_tight():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    cone = rep.create.cone(semantics=[("prim", "cone")], position=(100, 0, 0))
    sphere = rep.create.sphere(semantics=[("prim", "sphere")], position=(-100, 0, 0))
    invalid_type = rep.create.cube(semantics=[("shape", "boxy")], position=(0, 100, 0))

    # Setup semantic filter
    sd.SyntheticData.Get().set_instance_mapping_semantic_filter("prim:*")

    cam = rep.create.camera(position=(500,500,500), look_at=cone)
    rp = rep.create.render_product(cam, (1024, 512))

    bbox_2d_tight = rep.AnnotatorRegistry.get_annotator("bounding_box_2d_tight")
    bbox_2d_tight.attach(rp)

    await rep.orchestrator.step_async()
    data = bbox_2d_tight.get_data()
    print(data)
    # {
    #   'data': array([
    #       (0, 443, 198, 581, 357, 0.0),
    #       (1, 245,  94, 368, 220, 0.3823),
    #       dtype=[('semanticId', '<u4'),
    #              ('x_min', '<i4'),
    #              ('y_min', '<i4'),
    #              ('x_max', '<i4'),
    #              ('y_max', '<i4'),
    #              ('occlusionRatio', '<f4')]),
    #   'info': {
    #       'bboxIds': array([0, 1], dtype=uint32),
    #       'idToLabels': {'0': {'prim': 'cone'}, '1': {'prim': 'sphere'}},
    #       'primPaths': ['/Replicator/Cone_Xform', '/Replicator/Sphere_Xform']}
    #   }
    # }

import asyncio
asyncio.ensure_future(test_bbox_2d_tight())

Bounding Box 3D

Annotator Name: bounding_box_3d

Outputs 3D bounding box of each entity with semantics in the camera’s viewport.

Output Format

The bounding box annotator returns a dictionary with the bounds and semantic id found under the data key, while other information is under the info key: idToLabels, bboxIds and primPaths.

{
    "data": np.dtype(
                [
                    ("semanticId", "<u4"),          # Semantic identifier which can be transformed into a readable label using the `idToLabels` mapping
                    ("x_min", "<i4"),               # Minimum bound in x axis in local reference frame (in world units)
                    ("y_min", "<i4"),               # Minimum bound in y axis in local reference frame (in world units)
                    ("x_max", "<i4"),               # Maximum bound in x axis in local reference frame (in world units)
                    ("y_max", "<i4"),               # Maximum bound in y axis in local reference frame (in world units)
                    ("z_min", "<i4"),               # Minimum bound in z axis in local reference frame (in world units)
                    ("z_max", "<i4"),               # Maximum bound in z axis in local reference frame (in world units)
                    ("transform", "<i4"),           # Local to world transformation matrix (transforms the bounds from local frame to world frame)
                    ('occlusionRatio', '<f4')]),    # Occlusion (visible pixels / total pixels), where `0.0` is fully visible and `1.0` is fully occluded. See additional notes below.
                ],
    "info": {
        "idToLabels": {<semanticId>: <semantic_labels>},    # mapping from integer semantic ID to a comma delimited list of associated semantics
        "bboxIds": [<bbox_id_0>, ..., <bbox_id_n>],         # ID specific to bounding box annotators allowing easy mapping between different bounding box annotators.
        "primPaths": [<prim_path_0>, ... <prim_path_n>],    # prim path tied to each bounding box
    }
}

Note

bounding_box_3d are generated regardless of occlusion. bounding box dimensions (<axis>_min, <axis>_max) are expressed in stage units. bounding box transform is expressed in the world reference frame.

occlusionRatio can only provide valid values for prims composed of a single mesh. Multi-mesh labelled prims will return a value of -1 indicating that no occlusion value is available.

Example

import omni.replicator.core as rep
import omni.syntheticdata as sd

async def test_bbox_3d():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    cone = rep.create.cone(semantics=[("prim", "cone")], position=(100, 0, 0))
    sphere = rep.create.sphere(semantics=[("prim", "sphere")], position=(-100, 0, 0))
    invalid_type = rep.create.cube(semantics=[("shape", "boxy")], position=(0, 100, 0))

    # Setup semantic filter
    sd.SyntheticData.Get().set_instance_mapping_semantic_filter("prim:*")

    cam = rep.create.camera(position=(500,500,500), look_at=cone)
    rp = rep.create.render_product(cam, (1024, 512))

    bbox_3d = rep.AnnotatorRegistry.get_annotator("bounding_box_3d")
    bbox_3d.attach(rp)

    await rep.orchestrator.step_async()
    data = bbox_3d.get_data()
    print(data)
    # {
    #   'data': array([
    #       (0, -50., -50., -50., 50., 49.9999, 50., [[   1.,    0.,    0.,    0.], [   0.,    1.,    0.,    0.], [   0.,    0.,    1.,    0.], [ 100.,    0.,    0.,    1.]]),
    #       (1, -50., -50., -50., 50., 50.    , 50., [[   1.,    0.,    0.,    0.], [   0.,    1.,    0.,    0.], [   0.,    0.,    1.,    0.], [-100.,    0.,    0.,    1.]]),
    #       dtype=[('semanticId', '<u4'),
    #              ('x_min', '<i4'),
    #              ('y_min', '<i4'),
    #              ('x_max', '<i4'),
    #              ('y_max', '<i4'),
    #              ("z_min", "<i4"),
    #              ("z_max", "<i4"),
    #              ("transform", "<i4")]),
    #   'info': {
    #       'bboxIds': array([0, 1], dtype=uint32),
    #       'idToLabels': {'0': {'prim': 'cone'}, '1': {'prim': 'sphere'}},
    #       'primPaths': ['/Replicator/Cone_Xform_03', '/Replicator/Sphere_Xform_03']}
    #   }
    # }

import asyncio
asyncio.ensure_future(test_bbox_3d())

Distance to Camera

Annotator Name: distance_to_camera

Outputs a depth map from objects to camera positions. The distance_to_camera annotator produces a 2d array of types np.float32 with 1 channel.

Data Details

The unit for distance to camera is in meters (For example, if the object is 1000 units from the camera, and the meters_per_unit variable of the scene is 100, the distance to camera would be 10).
0 in the 2d array represents infinity (which means there is no object in that pixel).

Distance to Image Plane

Annotator Name: distance_to_image_plane

Outputs a depth map from objects to image plane of the camera. The distance_to_image_plane annotator produces a 2d array of types np.float32 with 1 channel.

Data Details

The unit for distance to image plane is in meters (For example, if the object is 1000 units from the image plane of the camera, and the meters_per_unit variable of the scene is 100, the distance to camera would be 10).
0 in the 2d array represents infinity (which means there is no object in that pixel).

Semantic Segmentation

Annotator Name: semantic_segmentation

Outputs semantic segmentation of each entity in the camera’s viewport that has semantic labels.

Initialization Parameters

colorize (bool): whether to output colorized semantic segmentation or non-colorized one.

Output Format

Semantic segmentation image:

If colorize is set to true, the image will be a 2d array of types np.uint8 with 4 channels.
Different colors represent different semantic labels.
If colorize is set to false, the image will be a 2d array of types np.uint32 with 1 channel, which is the semantic id of the entities.

ID to labels json file:

If colorize is set to true, it will be the mapping from color to semantic labels.
If colorize is set to false, it will be the mapping from semantic id to semantic labels.

Note

The semantic labels of an entity will be the semantic labels of itself, plus all the semantic labels it inherit from its parent, and semantic labels with same type will be concatenated, separated by comma.

For example, if an entity has a semantic label of [{“class”: “cube”}], and its parent has [{“class”: “rectangle”}]. Then the final semantic labels of that entity will be [{“class”: “rectangle, cube”}].

Instance ID Segmentation

Annotator Name: instance_id_segmentation

Outputs instance id segmentation of each entity in the camera’s viewport. The instance id is unique for each prim in the scene with different paths.

Initialization Parameters

colorize (bool): whether to output colorized semantic segmentation or non-colorized one.

Output Format

Instance ID segmentation image:

If colorize is set to true, the image will be a 2d array of types np.uint8 with 4 channels.
Different colors represent different instance ids.
If colorize is set to false, the image will be a 2d array of types np.uint32 with 1 channel, which is the instance id of the entities.

ID to labels json file:

If colorize is set to true, it will be the mapping from color to usd prim path of that entity.
If colorize is set to false, it will be the mapping from instance id to usd prim path of that entity.

Note

The instance id is assigned in a way that each of the leaf prim in the scene will be assigned to an instance id, no matter if it has semantic labels or not.

Instance Segmentation

Annotator Name: instance_segmentation

Outputs instance segmentation of each entity in the camera’s viewport. The main difference between instance id segmentation and instance segmentation are that instance segmentation annotator goes down the hierarchy to the lowest level prim which has semantic labels, which instance id segmentation always goes down to the leaf prim.

Initialization Parameters

colorize (bool): whether to output colorized instance segmentation or non-colorized one.

Output Format

Instance segmentation image:

If colorize is set to true, the image will be a 2d array of types np.uint8 with 4 channels.
Different colors represent different semantic instances.
If colorize is set to false, the image will be a 2d array of types np.uint32 with 1 channel, which is the instance id of the semantic entities.

ID to labels json file:

If colorize is set to true, it will be the mapping from color to usd prim path of that semantic entity.
If colorize is set to false, it will be the mapping from instance id to usd prim path of that semantic entity.

ID to semantic json file:

If colorize is set to true, it will be the mapping from color to semantic labels of that semantic entity.
If colorize is set to false, it will be the mapping from instance id to semantic labels of that semantic entity.

Note

Two prims with same semantic labels but live in different USD path will have different ids. If two prims have no semantic labels, and they have a same parent which has semantic labels, they will be classified as the same instance.

Point Cloud

Annotator Name: pointcloud

Outputs a 2D array of shape (N, 3) representing the points sampled on the surface of the prims in the viewport, where N is the number of point.

Output Format

The pointcloud annotator returns positions of the points found under the data key. Additional information is under the info key: pointRgb, pointNormals and pointSemantic.

{
    "data": np.dtype(np.float32),  # position value of each point of shape (N, 3)
    "info": {
        'pointNormals': ndarray[...]    #  shape=(<num_points> * 4), dtype=float32),
        'pointRgb': ndarray[...],       #  shape=(<num_points> * 4), dtype=uint8),
        'pointSemantic': dnarray[...],  #  shape=(<num_points>), dtype=uint8),
    }
}

Data Details

Point positions are in the world space.
Sample resolution is determined by the resolution of the render product.

Note

To get the mapping from semantic id to semantic labels, pointcloud annotator is better used with semantic segmentation annotator, and users can extract the idToLabels data from the semantic segmentation annotator.

Example 1

The Pointcloud annotator captures prims seen in the camera, and samples the points on the surface of the prims, based on the resolution of the render product attached to the camera.

Additional to the points sampled, it also outputs rgb, normals, and semantic id values associated to the prim where that point belongs to.

For prims without any valid semantic labels, pointcloud annotator will ignore it.

import asyncio
import omni.replicator.core as rep

async def test_pointcloud():
    # Add Default Light
    distance_light = rep.create.light(rotation=(315,0,0), intensity=3000, light_type="distant")

    # Pointcloud only capture prims with valid semantics
    W, H = (1024, 512)
    cube = rep.create.cube(position=(0, 0, 0), semantics=[("class", "cube")])
    camera = rep.create.camera(position=(200., 200., 200.), look_at=cube)
    render_product = rep.create.render_product(camera, (W, H))

    pointcloud_anno = rep.annotators.get("pointcloud")
    pointcloud_anno.attach(render_product)

    await rep.orchestrator.step_async()

    pc_data = pointcloud_anno.get_data()
    print(pc_data)
    # {
    #     'data': array([...], shape=(<num_points>, 3), dtype=float32),
    #     'info': {
    #         'pointNormals': [ 0.000e+00 1.00e+00 -1.5259022e-05 ... 0.00e+00 -1.5259022e-05 1.00e+00], shape=(<num_points> * 4), dtype=float32),
    #         'pointRgb': [241 240 241 ... 11  12 255], shape=(<num_points> * 4), dtype=uint8),
    #         'pointSemantic': [2 2 2 ... 2 2 2], shape=(<num_points>), dtype=uint8),
    #
    #     }
    # }

asyncio.ensure_future(test_pointcloud())

Example 2

In this example, we demonstrate a scenario where multiple camera captures are taken to produce a more complete pointcloud, utilizing the excellent open3d library to export a colored ply file.

import os
import asyncio
import omni.replicator.core as rep
import open3d as o3d
import numpy as np

async def test_pointcloud():
    # Pointcloud only capture prims with valid semantics
    cube = rep.create.cube(semantics=[("class", "cube")])
    camera = rep.create.camera()
    render_product = rep.create.render_product(camera, (1024, 512))

    pointcloud_anno = rep.annotators.get("pointcloud")
    pointcloud_anno.attach(render_product)

    # Camera positions to capture the cube
    camera_positions = [(500, 500, 0), (-500, -500, 0), (500, 0, 500), (-500, 0, -500)]

    with rep.trigger.on_frame(num_frames=len(camera_positions)):
        with camera:
            rep.modify.pose(position=rep.distribution.sequence(camera_positions), look_at=cube)  # make the camera look at the cube

    # Accumulate points
    points = []
    points_rgb = []
    for _ in range(len(camera_positions)):
        await rep.orchestrator.step_async()

        pc_data = pointcloud_anno.get_data()
        points.append(pc_data["data"])
        points_rgb.append(pc_data["info"]["pointRgb"].reshape(-1, 4)[:, :3])

    # Output pointcloud as .ply file
    ply_out_dir = os.path.join(os.path.dirname(os.path.realpath(__file__)), "out")
    os.makedirs(ply_out_dir, exist_ok=True)

    pc_data = np.concatenate(points)
    pc_rgb = np.concatenate(points_rgb)

    pcd = o3d.geometry.PointCloud()
    pcd.points = o3d.utility.Vector3dVector(pc_data)
    pcd.colors = o3d.utility.Vector3dVector(pc_rgb)
    o3d.io.write_point_cloud(os.path.join(ply_out_dir, "pointcloud.ply"), pcd)

asyncio.ensure_future(test_pointcloud())

Skeleton Data

Annotator Name: skeleton_data

The skeleton data annotator outputs pose information about the skeletons in the scene view.

Output Format

Parameter	Data Type	Description
animationVariant	list(<num_skeletons>, dtype=str)	Animation variant name for each skeleton
assetPath	list(<num_skeletons>, dtype=str)	Asset path for each skeleton
globalTranslations	array((<num_joints>, 3), dtype=float32)	Global translation of each joint
globalTranslationsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
inView	array(<num_skeletons>, dtype=bool)	If the skeleton is in view of the camera
jointOcclusions	array(<num_joints>, dtype=bool)	For each joint, True if joint is occluded, otherwise False
jointOcclusionsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
localRotations	array((<num_joints>, 4), dtype=float32)	Local rotation of each joint
localRotationsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
numSkeletons	<num_skeletons>	Number of skeletons in scene
occlusionTypes	list(str)	For each joint, the type of occlusion
occlusionTypesSizes	array([<num_skeletons>], dtype=int32)	Size of each set of joints per skeleton
restGlobalTranslations	array((<num_joints>, 3), dtype=float32)	Global translation for each joint at rest
restGlobalTranslationsSizes	array([<num_skeletons>], dtype=int32)	Size of each set of joints per skeleton
restLocalRotations	array((<num_joints>, 4), dtype=float32)	Local rotation of each join at rest
restLocalRotationsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
restLocalTranslations	array((<num_joints>, 3), dtype=float32)	Local translation of each join at rest
restLocalTranslationsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
skeletonJoints	list(str)	List of skeleton joints, encoded as a string
skeletonParents	array(<num_joints>, dtype=int32)	Which joint is the parent of the index, -1 is root
skeletonParentsSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton
skelName	list(<num_skeletons>, dtype=str)	Name of each skeleton
skelPath	list(<num_skeletons>, dtype=str)	Path of each skeleton prim
translations2d	array((<num_joints>, 2), dtype=float32)	Screen space joint position in pixels
translations2dSizes	array(<num_skeletons>, dtype=int32)	Size of each set of joints per skeleton

This annotator returns additional data as a single string held in a dictionary with the key skeleton_data for backwards compatibility with the original implementation of this annotator. Use eval(data[“skeleton_data”]) to extract the attributes from this string.

Example

Below is an example script that outputs 10 images with skeleton pose annotation.

import asyncio
import omni.replicator.core as rep

# Define paths for the character
PERSON_SRC = 'omniverse://localhost/NVIDIA/Assets/Characters/Reallusion/Worker/Worker.usd'

async def test_skeleton_data():
    # Human Model
    person = rep.create.from_usd(PERSON_SRC, semantics=[('class', 'person')])
    # Area to scatter cubes in
    area = rep.create.cube(scale=2, position=(0.0, 0.0, 100.0), visible=False)

    # Create the camera and render product
    camera = rep.create.camera(position=(25, -421.0, 182.0), rotation=(77.0, 0.0, 3.5))
    render_product = rep.create.render_product(camera, (1024, 1024))

    def randomize_spheres():
        spheres = rep.create.sphere(scale=0.1, count=100)
        with spheres:
            rep.randomizer.scatter_3d(area)
        return spheres.node

    rep.randomizer.register(randomize_spheres)

    with rep.trigger.on_frame(interval=10, max_execs=5):
        rep.randomizer.randomize_spheres()

    # Attach annotator
    skeleton_anno = rep.annotators.get("skeleton_data")
    skeleton_anno.attach(render_product)

    await rep.orchestrator.step_async()

    data = skeleton_anno.get_data()
    print(data)
    # {
    #     'animationVariant': ['None'],
    #     'assetPath': ['Bones/Worker.StandingDiscussion_LookingDown_M.usd'],
    #     'globalTranslations': array([[  0.  ,   0.  ,   0.  ], ..., [-21.64,   2.58, 129.8 ]], dtype=float32),
    #     'globalTranslationsSizes': array([101], dtype=int32),
    #     'inView': array([ True]),
    #     'jointOcclusions': array([ True, False, ..., False, False]),
    #     'jointOcclusionsSizes': array([101], dtype=int32),
    #     'localRotations': array([[ 1.  ,  0.  ,  0.  ,  0.  ], ..., [ 1.  ,  0.  , -0.09, -0.  ]], dtype=float32),
    #     'localRotationsSizes': array([101], dtype=int32),
    #     'numSkeletons': 1,
    #     'occlusionTypes': ["['BACKGROUND', 'None', ..., 'None', 'None']"],
    #     'occlusionTypesSizes': array([101], dtype=int32),
    #     'restGlobalTranslations': array([[  0.  ,   0.  ,   0.  ], ..., [-31.86,   8.96, 147.72]], dtype=float32),
    #     'restGlobalTranslationsSizes': array([101], dtype=int32),
    #     'restLocalRotations': array([[ 1.  ,  0.  ,  0.  ,  0.  ], ..., [ 1.  ,  0.  ,  0.  , -0.  ]], dtype=float32),
    #     'restLocalRotationsSizes': array([101], dtype=int32),
    #     'restLocalTranslations': array([[  0.  ,   0.  ,   0.  ], ..., [ -0.  ,  12.92,   0.01]], dtype=float32),
    #     'restLocalTranslationsSizes': array([101], dtype=int32),
    #     'skeletonJoints': ["['RL_BoneRoot', 'RL_BoneRoot/Hip', ..., 'RL_BoneRoot/Hip/Waist/Spine01/Spine02/R_Clavicle/R_Upperarm/R_UpperarmTwist01/R_UpperarmTwist02']"],
    #     'skeletonParents': array([-1,  0,  1, ..., 97, 78, 99], dtype=int32),
    #     'skeletonParentsSizes': array([101], dtype=int32),
    #     'skelName': ['Worker'],
    #     'skelPath': ['/Replicator/Ref_Xform/Ref/ManRoot/Worker/Worker'],
    #     'translations2d': array([[513.94, 726.03],
    #                              [514.42, 480.42],
    #                              [514.42, 480.42],
    #                              ...,
    #                              [499.45, 450.9 ],
    #                              [466.3 , 354.6 ],
    #                              [455.09, 388.56]], dtype=float32),
    #    'translations2dSizes': array([101], dtype=int32),
    #    'skeletonData': ...   # string data representation for backward compatibility
    # }

asyncio.ensure_future(test_skeleton_data())

Motion Vectors

Annotator Name: motion_vectors

Outputs a 2D array of motion vectors representing the relative motion of a pixel in the camera’s viewport between frames.

The MotionVectors annotator returns the per-pixel motion vectors in in image space.

Output Format

array((height, width, 4), dtype=<np.float32>)

The components of each entry in the 2D array represent four different values encoded as floating point values:

x: motion distance in the horizontal axis (image width) with movement to the left of the image being positive and movement to the right being negative.
y: motion distance in the vertical axis (image height) with movement towards the top of the image being positive and movement to the bottom being negative.
z: unused
w: unused

Example

import asyncio
import omni.replicator.core as rep

async def test_motion_vectors():
    # Add an object to look at
    cone = rep.create.cone()

    # Add motion to object
    cone_prim = cone.get_output_prims()["prims"][0]
    cone_prim.GetAttribute("xformOp:translate").Set((-100, 0, 0), time=0.0)
    cone_prim.GetAttribute("xformOp:translate").Set((100, 50, 0), time=10.0)

    camera = rep.create.camera()
    render_product = rep.create.render_product(camera, (512, 512))

    motion_vectors_anno = rep.annotators.get("MotionVectors")
    motion_vectors_anno.attach(render_product)

    # Take a step to render the initial state (no movement yet)
    await rep.orchestrator.step_async()

    # Capture second frame (now the timeline is playing)
    await rep.orchestrator.step_async()
    data = motion_vectors_anno.get_data()
    print(data.shape, data.dtype, data.reshape(-1, 4).min(axis=0), data.reshape(-1, 4).max(axis=0))
    # (1024, 512, 4), float32,  [-93.80073  -1.       -1.       -1.     ] [ 0.       23.450201  1.        1.      ]

asyncio.ensure_future(test_motion_vectors())

Note

The values represent motion relative to camera space.

Cross Correspondence

Annotator Name: cross_correspondence

The cross correspondence annotator outputs a 2D array representing the camera optical flow map of the camera’s viewport against a reference viewport.

To enable the cross correspondence annotation, the camera attached to the render product annotated with cross correspondence must have the attribute crossCameraReferenceName set to the (unique) name (not path) of a second camera (itself attached to a second render product). The Projection Type of the two cameras needs to be of type fisheyePolynomial (Camera –> Fisheye Lens –> Projection Type –> fisheyePolynomial).

Output Format

The Cross Correspondence annotator produces the cross correspondence between pixels seen from two cameras.

The components of each entry in the 2D array represent four different values encoded as floating point values:

x: dx - difference to the x value of of the corresponding pixel in the reference viewport. This value is normalized to [-1.0, 1.0]
y: dy - difference to the y value of of the corresponding pixel in the reference viewport. This value is normalized to [-1.0, 1.0]
z: occlusion mask - boolean signifying that the pixel is occluded or truncated in one of the cross referenced viewports. Floating point value represents a boolean (1.0 = True, 0.0 = False)
w: geometric occlusion calculated - boolean signifying that the pixel can or cannot be tested as having occluded geometry (e.g. no occlusion testing is performed on missed rays) (1.0 = True, 0.0 = False)

array((height, width, 4), dtype=<np.float32>)

Example

import asyncio
import omni.replicator.core as rep
from pxr import Sdf

async def test_cross_correspondence():
    # Add an object to look at
    rep.create.cone()

    # Add stereo camera pair
    stereo = rep.create.stereo_camera(position=(20, 0, 300), projection_type="fisheye_polynomial", stereo_baseline=20)

    # Add cross correspondence attribute
    stereo_L_prim = stereo.get_output_prims()["prims"][0].GetChildren()[0].GetChildren()[0]
    stereo_L_prim.CreateAttribute("crossCameraReferenceName", Sdf.ValueTypeNames.String)

    # Set attribute to refer to second camera name - beware of scenes with multiple cameras that share the same name!
    stereo_L_prim.GetAttribute("crossCameraReferenceName").Set("StereoCam_R")

    render_products = rep.create.render_product(stereo, (512, 512))

    # Add annotator to left render product
    anno = rep.annotators.get("cross_correspondence")
    anno.attach(render_products[0])

    await rep.orchestrator.step_async()
    data = anno.get_data()
    print(data.shape, data.dtype)
    # (512, 512, 4), float32

asyncio.ensure_future(test_cross_correspondence())

Note

Both cameras must have the cameraProjectionType attribute set to fisheyePolynomial

The annotated camera must have the crossCameraReferenceName attribute set to the name of the second camera

To avoid unexpected results, ensure that the referenced camera has a unique name

Camera Parameters

Annotator Name: camera_params

The Camera Parameters annotator returns the camera details for the camera corresponding to the render product to which the annotator is attached.

Data Details

Parameter	Description
cameraFocalLength	Camera focal length
cameraFocusDistance	Camera focus distance
cameraFStop	Camera fStop value
cameraAperture	Camera horizontal and vertical aperture
cameraApertureOffset	Camera horizontal and vertical aperture offset
renderProductResolution	RenderProduct resolution
cameraModel	Camera model name
cameraViewTransform	Camera to world transformation matrix
cameraProjection	Camera projection matrix
cameraFisheyeNominalWidth	Camera fisheye nominal width
cameraFisheyeNominalHeight	Camera fisheye nominal height
cameraFisheyeOpticalCentre	Camera fisheye optical centre
cameraFisheyeMaxFOV	Camera fisheye maximum field of view
cameraFisheyePolynomial	Camera fisheye polynomial
cameraNearFar	Camera near/far clipping range

Example

import asyncio
import omni.replicator.core as rep

async def test_camera_params():
    camera_1 = rep.create.camera()
    camera_2 = rep.create.camera(
        position=(100, 0, 0),
        projection_type="fisheye_polynomial"
    )

    render_product_1 = rep.create.render_product(camera, (1024, 1024))
    render_product_2 = rep.create.render_product(camera, (1024, 1024))


    anno_1 = rep.annotators.get("CameraParams").attach(render_product_1)
    anno_2 = rep.annotators.get("CameraParams").attach(render_product_2)

    await rep.orchestrator.step_async()

    print(anno_1)
    # {'cameraAperture': array([20.955 , 15.2908], dtype=float32),
    #     'cameraApertureOffset': array([0., 0.], dtype=float32),
    #     'cameraFisheyeMaxFOV': 0.0,
    #     'cameraFisheyeNominalHeight': 0,
    #     'cameraFisheyeNominalWidth': 0,
    #     'cameraFisheyeOpticalCentre': array([0., 0.], dtype=float32),
    #     'cameraFisheyePolynomial': array([0., 0., 0., 0., 0.], dtype=float32),
    #     'cameraFocalLength': 24.0,
    #     'cameraFocusDistance': 400.0,
    #     'cameraFStop': 0.0,
    #     'cameraModel': 'pinhole',
    #     'cameraNearFar': array([1., 1000000.], dtype=float32),
    #     'cameraProjection': array([ 2.291,  0.   ,  0.   ,  0.   ,
    #                                 1.   ,  2.291,  0.   ,  0.   ,
    #                                 1.   ,  0.   ,  0.   , -1.   ,
    #                                 1.   ,  0.   ,  1.   ,  0.   ]),
    #     'cameraViewTransform': array([1., 0., 0., 0., 0., 1., 0., 0., 0., 0., 1., 0., 0., 0., 0., 1.]),
    #     'metersPerSceneUnit': 0.009999999776482582,
    #     'renderProductResolution': array([1024, 1024], dtype=int32)
    # }

    print(anno_2)
    # {
    #     'cameraAperture': array([20.955 , 15.291], dtype=float32),
    #     'cameraApertureOffset': array([0., 0.], dtype=float32),
    #     'cameraFisheyeMaxFOV': 200.0,
    #     'cameraFisheyeNominalHeight': 1216,
    #     'cameraFisheyeNominalWidth': 1936,
    #     'cameraFisheyeOpticalCentre': array([970.9424, 600.375 ], dtype=float32),
    #     'cameraFisheyePolynomial': array([0.     , 0.002, 0.     , 0.     , 0.     ], dtype=float32),
    #     'cameraFocalLength': 24.0,
    #     'cameraFocusDistance': 400.0,
    #     'cameraFStop': 0.0,
    #     'cameraModel': 'fisheyePolynomial',
    #     'cameraNearFar': array([1., 1000000.], dtype=float32),
    #     'cameraProjection': array([ 2.291,  0.   ,  0.   ,  0.   ,
    #                                 1.   ,  2.291,  0.   ,  0.   ,
    #                                 2.   ,  0.   ,  0.   , -1.   ,
    #                                 3.   ,  0.   ,  1.   ,  0.   ]),
    #     'cameraViewTransform': array([1., 0., 0., 0., 0., 1., 0., 0., 0., 0., 1., 0., -100., 0., 0., 1.]),
    #   'metersPerSceneUnit': 0.009999999776482582,
    #   'renderProductResolution': array([1024, 1024], dtype=int32)
    #  }

asyncio.ensure_future(test_camera_params())

Reference Time

Outputs the reference time corresponding to the render and associated annotations.

Output Format

The reference time annotator returns a numerator and denominator representing the time corresponding to the render and associated annotations.

{
    'referenceTimeNumerator': int,
    'referenceTimeDenominator': int,
}

Example

import asyncio
import omni.replicator.core as rep

async def test_reference_time():
    W, H = (1024, 512)
    camera = rep.create.camera()
    render_product = rep.create.render_product(camera, (W, H))

    ref_time_anno = rep.annotators.get("ReferenceTime")
    ref_time_anno.attach(render_product)

    await rep.orchestrator.step_async()

    ref_time_data = ref_time_anno.get_data()
    print(ref_time_data)
    # {
    #   'referenceTimeNumerator': <numerator>,
    #   'referenceTimeDenominator': <denominator>,
    # }

asyncio.ensure_future(test_reference_time())

RT Annotators

RT Annotators are only available in RayTracedLighting rendering mode (RTX - Real-Time)

Example

import asyncio
import omni.replicator.core as rep

async def test_pt_anno():
    # Set rendermode to PathTracing
    rep.settings.set_render_rtx_realtime()

    # Create an interesting scene
    red_diffuse = rep.create.material_omnipbr(diffuse=(1, 0, 0.2), roughness=1.0)
    metallic_reflective = rep.create.material_omnipbr(roughness=0.01, metallic=1.0)
    glow = rep.create.material_omnipbr(emissive_color=(1.0, 0.5, 0.4), emissive_intensity=100000.0)
    rep.create.cone(material=metallic_reflective)
    rep.create.cube(position=(100, 50, -100), material=red_diffuse)
    rep.create.sphere(position=(-100, 50, 100), material=glow)
    ground = rep.create.plane(scale=(100, 1, 100), position=(0, -50, 0))

    # Attach render product
    W, H = (1024, 512)
    camera = rep.create.camera(position=(400., 400., 400.), look_at=ground)
    render_product = rep.create.render_product(camera, (W, H))

    anno = rep.annotators.get("SmoothNormal")
    anno.attach(render_product)

    await rep.orchestrator.step_async()

    data = anno.get_data()
    print(data.shape, data.dtype)
    # (512, 1024, 4), float32

asyncio.ensure_future(test_pt_anno())

Tiled Sensor Annotator

The Tiled Sensor Annotator tiles multiple camera outputs into a tiled buffer, enabling faster data transfer. Currently, albedo RGB and depth (distance to camera) are supported as outputs.

You can view the output of the tiled sensor by setting the viewport to that sensor. Also, ensure the resolution of the viewport is the same as the resolution of the tiled sensor.

Example

import asyncio
import omni.replicator.core as rep

async def test_tiled_sensor():
    await omni.usd.get_context().new_stage_async()

    seed = 1234
    # Set up the scene
    cube = rep.create.cube(semantics=[("class", "cube")])
    cone = rep.create.cone(semantics=[("class", "cone")], position=(0, 80, 0))

    with cube:
        rep.randomizer.color(colors=rep.distribution.uniform((0, 0, 0), (1, 1, 1), seed=4321))

    with cone:
        rep.randomizer.color(colors=rep.distribution.uniform((0, 0, 0), (1, 1, 1), seed=seed))

    # create a list of camera looking at the scene
    cams = []
    for i in range(32):
        cam = rep.create.camera(
            position=rep.distribution.uniform((200, 200, 200), (500, 500, 500), seed=seed), look_at=cube
        )
        cams.append(cam)

    # create the tiled sensor that takes in the list of cameras as input
    # Notice that tiled_resolution must be divisible by camera_resolution in each dimension.
    # The example outputs RGB, but it can also outputs depth or both.
    sensor = rep.create.tiled_sensor(
        cameras=cams, camera_resolution=(128, 128), tiled_resolution=(1024, 512), output_types=["rgb"]
    )

    # tiled sensor is treated just like a camera. Create a render product that is tied to the sensor.
    rp = rep.create.render_product(camera=sensor, resolution=(1024, 512))

    # The name of the tiled sensor annotator is RtxSensorGpu and it resides on GPU.
    anno = rep.annotators.get("RtxSensorGpu")
    anno.attach(rp)

    await rep.orchestrator.step_async()

    # Post process the data
    sensor_data = anno.get_data().reshape(512, 1024, -1) * 255

asyncio.ensure_future(test_tiled_sensor())

Tiled Sensor RGB — The RGB output of the Tiled Sensor if run the code above

SmoothNormal

Output Format

np.ndtype(np.float32)   # shape: (H, W, 4)

BumpNormal

Output Format

np.ndtype(np.float32)   # shape: (H, W, 4)

AmbientOcclusion

Output Format

np.ndtype(np.float16)   # shape: (H, W, 4)

Motion2d

Output Format

np.ndtype(np.float32)   # shape: (H, W, 4)

DiffuseAlbedo

Output Format

np.ndtype(np.uint8) # shape: (H, W, 4)

SpecularAlbedo

Output Format

np.ndtype(np.float16)   # shape: (H, W, 4)

Roughness

Output Format

np.ndtype(np.uint8) # shape: (H, W, 4)

DirectDiffuse

Output Format

np.ndtype(np.float16)   # shape: (H, W, 4)

DirectSpecular

Output Format

np.ndtype(np.float16)   # shape: (H, W, 4)

Reflections

Output Format

np.ndtype(np.float32)   # shape: (H, W, 4)

IndirectDiffuse

Output Format

np.ndtype(np.float16)   # shape: (H, W, 4)

DepthLinearized

Output Format

np.ndtype(np.float32)   # shape: (H, W, 1)

EmissionAndForegroundMask

Output Format

np.ndtype(np.float16)   # shape: (H, W, 1)

PathTracing Annotators

PathTracing Annotators are only available in PathTracing rendering mode (RTX - Interactive). In addition, the following carb settings must be set on app launch:

rtx-transient.aov.enableRtxAovs = true
rtx-transient.aov.enableRtxAovsSecondary = true

Example

import asyncio
import omni.replicator.core as rep

async def test_pt_anno():
    # Set rendermode to PathTracing
    rep.settings.set_render_pathtraced()

    # Create an interesting scene
    red_diffuse = rep.create.material_omnipbr(diffuse=(1, 0, 0.2), roughness=1.0)
    metallic_reflective = rep.create.material_omnipbr(roughness=0.01, metallic=1.0)
    glow = rep.create.material_omnipbr(emissive_color=(1.0, 0.5, 0.4), emissive_intensity=100000.0)
    rep.create.cone(material=metallic_reflective)
    rep.create.cube(position=(100, 50, -100), material=red_diffuse)
    rep.create.sphere(position=(-100, 50, 100), material=glow)
    ground = rep.create.plane(scale=(100, 1, 100), position=(0, -50, 0))

    # Attach render product
    W, H = (1024, 512)
    camera = rep.create.camera(position=(400., 400., 400.), look_at=ground)
    render_product = rep.create.render_product(camera, (W, H))

    anno = rep.annotators.get("PtGlobalIllumination")
    anno.attach(render_product)

    await rep.orchestrator.step_async()

    data = anno.get_data()
    print(data.shape, data.dtype)
    # (512, 1024, 4), float16

asyncio.ensure_future(test_pt_anno())

PtDirectIllumation

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtGlobalIllumination

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtReflections

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtRefractions

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtSelfIllumination

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtBackground

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtWorldNormal

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtWorldPos

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtZDepth

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtVolumes

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtDiffuseFilter

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtReflectionFilter

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtRefractionFilter

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte0

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte1

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte2

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte3

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte4

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte5

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte6

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)

PtMultiMatte7

Output Format

np.ndtype(np.float16)   # Shape: (Height, Width, 4)