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:
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 typesnp.uint8
with 4 channels.Different colors represent different semantic labels.
If
colorize
is set to false, the image will be a 2d array of typesnp.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 typesnp.uint8
with 4 channels.Different colors represent different instance ids.
If
colorize
is set to false, the image will be a 2d array of typesnp.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 typesnp.uint8
with 4 channels.Different colors represent different semantic instances.
If
colorize
is set to false, the image will be a 2d array of typesnp.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 coloured 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 correspondance annotation, the camera attached to the render product annotated with cross correspondance 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())
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)