CameraInfo does not follow render product resolution in Isaac Sim 5.1 (intended or bug?)

[DonghyunSung-MS]

Hi,

I observed an inconsistency between Image and CameraInfo when publishing camera data in Isaac Sim 5.1.

Setup

• USD camera resolution: 2448 x 2048
• Render product resolution: 1224 x 1024

Observed behavior

• /image topic:

  • width = 1224
  • height = 1024

• /camera_info topic:

  • width = 2448
  • height = 2048

So the image follows the render product resolution, but CameraInfo still reflects the original USD camera resolution.

Investigation

Looking into the source code of read_camera_info(render_product_path):

• It retrieves intrinsics (fx, fy, cx, cy) and imageSize from the camera prim (USD)
• It does not consider render product resolution

Therefore, CameraInfo is not scaled when render resolution is changed.

Minimal reproducible example

from isaacsim.sensors.camera import Camera
import numpy as np

prim_path = "/zivid/camera/camera_sensor"
resolution = (1224, 1024)

camera = Camera(
    prim_path=prim_path,
    resolution=resolution,
)

camera.initialize()

render_product = camera._render_product_path	

camera_info, _ = read_camera_info(render_product_path=render_product)

print("Render resolution:", camera.get_resolution())
print("CameraInfo width:", camera_info.width)
print("CameraInfo height:", camera_info.height)

Render resolution: (1224, 1024)
CameraInfo width: 2448
CameraInfo height: 2048

Question

Is this behavior intended?

In ROS2 ecosystem, Image and CameraInfo are expected to match:

• Same width/height
• Intrinsics corresponding to the actual image

Otherwise, downstream nodes (point cloud, SLAM, rectification) may produce incorrect results.

Expected behavior

If render product resolution is changed, CameraInfo should:

• Match the new resolution
• Scale intrinsics accordingly

Workaround

Currently, we manually scale:

• fx, fy, cx, cy
• width, height

based on render resolution.

Additional question

Is there a recommended way to generate CameraInfo that matches render product resolution?

Thanks.

[vick-yu]

Hi @zfrancisco-zivid,

Thanks for the detailed report and minimal repro — this is clearly documented.

You're correct that read_camera_info() does not account for the render product resolution when the camera uses the opencvPinhole or opencvFisheye lens distortion model. In those code paths, width, height, fx,
fy, cx, cy are all read directly from the camera prim attributes (e.g., omni:lensdistortion:opencvPinho le:imageSize) without scaling to match the actual render product resolution.

Interestingly, the fallback code path (when no lens distortion model is set) does use get_resolution(render_product_p ath) — so the inconsistency is limited to the opencvPinhole and opencvFisheye paths.

We're checking with the team on whether this is intended behavior or a bug, and will follow up here once we have clarity. In the meantime, your workaround of manually scaling the intrinsics and resolution is the
right approach:

  render_w, render_h = camera.get_resolution()                                                                                                                                                                      
  scale_x = render_w / camera_info.width                                                                                                                                                                            
  scale_y = render_h / camera_info.height                                                                                                                                                                           
                                                                                                                                                                                                                    
  camera_info.width = render_w                                                                                                                                                                                      
  camera_info.height = render_h                                                                                                                                                                                     
  camera_info.k[0] *= scale_x   # fx                                                                                                                                                                                
  camera_info.k[2] *= scale_x   # cx                                                                                                                                                                                
  camera_info.k[4] *= scale_y   # fy                                                                                                                                                                                
  camera_info.k[5] *= scale_y   # cy                                                                                                                                                                                
  camera_info.p[0] *= scale_x   # fx                                                                                                                                                                                
  camera_info.p[2] *= scale_x   # cx                                                                                                                                                                                
  camera_info.p[5] *= scale_y   # fy                                                                                                                                                                                
  camera_info.p[6] *= scale_y   # cy                                                                                                                                                                                

Will update this thread once we have more information.

[vick-yu]

An internal ticket has been created to tracked this issue with the opencvPinhole and opencvFisheye lens models and we will update this thread once we have clarity.

[vick-yu]

Hi @DonghyunSung-MS, great news — the bug you reported with read_camera_info() not scaling intrinsics and resolution for the opencvPinhole and opencvFisheye lens models has been fixed and will be included in Isaac Sim 6.0 (GA). Thank you for the thorough write-up, minimal repro, and workaround — it made tracking this down much easier. In the meantime, please continue using the manual scaling workaround. If you run into any further issues after upgrading, feel free to open a new discussion and reference this one: #545