Camera

Camera#

Camera sensors are uniquely defined by the use of the render_product, a structure for managing data generated by the rendering pipeline (images). Isaac Lab provides the ability to fully control how these renderings are created through camera parameters like focal length, pose, type, etc… and what kind of data you want to render through the use of Annotators, allowing you to record not only RGB, but also Instance segmentation, object pose, object ID, etc…

Rendered images are unique among the supported data types in Isaac Lab due to the inherently large bandwidth requirements for moving those data. A single 800 x 600 image with 32-bit color (a single float per pixel) clocks in at just under 2 MB. If we render at 60 fps and record every frame, that camera needs to move 120 MB/s. Multiply this by the number of cameras in an environment and environments in a simulation, and you can quickly see how scaling a naive vectorization of camera data could lead to bandwidth challenges. NVIDIA’s Isaac Lab leverages our expertise in GPU hardware to provide an API that specifically addresses these scaling challenges in the rendering pipeline.

Tiled Rendering#

Note

This feature is only available from Isaac Sim version 4.2.0 onwards.

Tiled rendering in combination with image processing networks require heavy memory resources, especially at larger resolutions. We recommend running 512 cameras in the scene on RTX 4090 GPUs or similar.

The Tiled Rendering APIs provide a vectorized interface for collecting data from camera sensors. This is useful for reinforcement learning environments where parallelization can be exploited to accelerate data collection and thus the training loop. Tiled rendering works by using a single render_product for all clones of a single camera in the scene. The desired dimensions of a single image and the number of environments are used to compute a much larger render_product, consisting of the tiled individual renders from the separate clones of the camera. When all cameras have populated their buffers the render product is “completed” and can be moved around as a single, large image, dramatically reducing the overhead for moving the data from the host to the device, for example. Only a single call is used to synchronize the device data, instead of one call per camera, and this is a big part of what makes the Tiled Rendering API more efficient for working with vision data.

Isaac Lab provides tiled rendering APIs for RGB, depth, along with other annotators through the TiledCamera class. Configurations for the tiled rendering APIs can be defined through the TiledCameraCfg class, specifying parameters such as the regex expression for all camera paths, the transform for the cameras, the desired data type, the type of cameras to add to the scene, and the camera resolution.

tiled_camera: TiledCameraCfg = TiledCameraCfg(
    prim_path="/World/envs/env_.*/Camera",
    offset=TiledCameraCfg.OffsetCfg(pos=(-7.0, 0.0, 3.0), rot=(0.9945, 0.0, 0.1045, 0.0), convention="world"),
    data_types=["rgb"],
    spawn=sim_utils.PinholeCameraCfg(
        focal_length=24.0, focus_distance=400.0, horizontal_aperture=20.955, clipping_range=(0.1, 20.0)
    ),
    width=80,
    height=80,
)

To access the tiled rendering interface, a TiledCamera object can be created and used to retrieve data from the cameras.

tiled_camera = TiledCamera(cfg.tiled_camera)
data_type = "rgb"
data = tiled_camera.data.output[data_type]

The returned data will be transformed into the shape (num_cameras, height, width, num_channels), which can be used directly as observation for reinforcement learning.

When working with rendering, make sure to add the --enable_cameras argument when launching the environment. For example:

python source/standalone/workflows/rl_games/train.py --task=Isaac-Cartpole-RGB-Camera-Direct-v0 --headless --enable_cameras

Annotators#

Both TiledCamera and Camera classes provide APIs for retrieving various types annotator data from replicator:

  • "rgb": A 3-channel rendered color image.

  • "rgba": A 4-channel rendered color image with alpha channel.

  • "distance_to_camera": An image containing the distance to camera optical center.

  • "distance_to_image_plane": An image containing distances of 3D points from camera plane along camera’s z-axis.

  • "depth": The same as "distance_to_image_plane".

  • "normals": An image containing the local surface normal vectors at each pixel.

  • "motion_vectors": An image containing the motion vector data at each pixel.

  • "semantic_segmentation": The semantic segmentation data.

  • "instance_segmentation_fast": The instance segmentation data.

  • "instance_id_segmentation_fast": The instance id segmentation data.

RGB and RGBA#

A scene captured in RGB

rgb data type returns a 3-channel RGB colored image of type torch.uint8, with dimension (B, H, W, 3).

rgba data type returns a 4-channel RGBA colored image of type torch.uint8, with dimension (B, H, W, 4).

To convert the torch.uint8 data to torch.float32, divide the buffer by 255.0 to obtain a torch.float32 buffer containing data from 0 to 1.

Depth and Distances#

A scene captured in RGB

distance_to_camera returns a single-channel depth image with distance to the camera optical center. The dimension for this annotator is (B, H, W, 1) and has type torch.float32.

distance_to_image_plane returns a single-channel depth image with distances of 3D points from the camera plane along the camera’s Z-axis. The dimension for this annotator is (B, H, W, 1) and has type torch.float32.

depth is provided as an alias for distance_to_image_plane and will return the same data as the distance_to_image_plane annotator, with dimension (B, H, W, 1) and type torch.float32.

Normals#

A scene captured in RGB

normals returns an image containing the local surface normal vectors at each pixel. The buffer has dimension (B, H, W, 3), containing the (x, y, z) information for each vector, and has data type torch.float32.

Motion Vectors#

motion_vectors returns the per-pixel motion vectors in image space, with a 2D array of motion vectors representing the relative motion of a pixel in the camera’s viewport between frames. The buffer has dimension (B, H, W, 2), representing x - the 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 and 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. The data type is torch.float32.

Semantic Segmentation#

A scene captured in RGB

semantic_segmentation outputs semantic segmentation of each entity in the camera’s viewport that has semantic labels. In addition to the image buffer, an info dictionary can be retrieved with tiled_camera.data.info['semantic_segmentation'] containing ID to labels information.

  • If colorize_semantic_segmentation=True in the camera config, a 4-channel RGBA image will be returned with dimension (B, H, W, 4) and type torch.uint8. The info idToLabels dictionary will be the mapping from color to semantic labels.

  • If colorize_semantic_segmentation=False, a buffer of dimension (B, H, W, 1) of type torch.int32 will be returned, containing the semantic ID of each pixel. The info idToLabels dictionary will be the mapping from semantic ID to semantic labels.

Instance ID Segmentation#

A scene captured in RGB

instance_id_segmentation_fast 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. In addition to the image buffer, an info dictionary can be retrieved with tiled_camera.data.info['instance_id_segmentation_fast'] containing ID to labels information.

The main difference between instance_id_segmentation_fast and instance_segmentation_fast are that instance segmentation annotator goes down the hierarchy to the lowest level prim which has semantic labels, where instance ID segmentation always goes down to the leaf prim.

  • If colorize_instance_id_segmentation=True in the camera config, a 4-channel RGBA image will be returned with dimension (B, H, W, 4) and type torch.uint8. The info idToLabels dictionary will be the mapping from color to USD prim path of that entity.

  • If colorize_instance_id_segmentation=False, a buffer of dimension (B, H, W, 1) of type torch.int32 will be returned, containing the instance ID of each pixel. The info idToLabels dictionary will be the mapping from instance ID to USD prim path of that entity.

Instance Segmentation#

A scene captured in RGB

instance_segmentation_fast outputs instance segmentation of each entity in the camera’s viewport. In addition to the image buffer, an info dictionary can be retrieved with tiled_camera.data.info['instance_segmentation_fast'] containing ID to labels and ID to semantic information.

  • If colorize_instance_segmentation=True in the camera config, a 4-channel RGBA image will be returned with dimension (B, H, W, 4) and type torch.uint8.

  • If colorize_instance_segmentation=False, a buffer of dimension (B, H, W, 1) of type torch.int32 will be returned, containing the instance ID of each pixel.

The info idToLabels dictionary will be the mapping from color to USD prim path of that semantic entity. The info idToSemantics dictionary will be the mapping from color to semantic labels of that semantic entity.

Current Limitations#

Due to current limitations in the renderer, we can have only one TiledCamera instance in the scene. For use cases that require a setup with more than one camera, we can imitate the multi-camera behavior by moving the location of the camera in between render calls in a step.

For example, in a stereo vision setup, the below snippet can be implemented:

# render image from "first" camera
camera_data_1 = self._tiled_camera.data.output["rgb"].clone() / 255.0
# update camera transform to the "second" camera location
self._tiled_camera.set_world_poses(
    positions=pos,
    orientations=rot,
    convention="world"
)
# step the renderer
self.sim.render()
self._tiled_camera.update(0, force_recompute=True)
# render image from "second" camera
camera_data_2 = self._tiled_camera.data.output["rgb"].clone() / 255.0

Note that this approach still limits the rendering resolution to be identical for all cameras. Currently, there is no workaround to achieve different resolution images using TiledCamera. The best approach is to use the largest resolution out of all of the desired resolutions and add additional scaling or cropping operations to the rendered output as a post-processing step.

In addition, there may be visible quality differences when comparing render outputs of different numbers of environments. Currently, any combined resolution that has a width less than 265 pixels or height less than 265 will automatically switch to the DLAA anti-aliasing mode, which does not perform up-sampling during anti-aliasing. For resolutions larger than 265 in both width and height dimensions, we default to using the “performance” DLSS mode for anti-aliasing for performance benefits. Anti-aliasing modes and other rendering parameters can be specified in the RenderCfg.