How to perform correct pose estimation in Isaac Sim using the tutorial's trained model

I am trying to perform pose estimation in IsaacSim using the trained model of ketchup used in the tutorial.

I imported the 3D model of the ketchup into the carter_warehouse_apriltags_worker.usd file and estimated from the camera data.
The model recognizes the ketchup bottle, but its pose is different than expected.
I have tried changing the resolution to 640*480 and changing the camera_matrix to be closer to the value described in the config.yaml file, but still get the same results.

Is there a way to get the pose estimation right?

Could you please provide some more information about your setup? A copy of your camera calibration file and the dope_config.yaml file would be helpful.

We are currently investigating a similar issue reported on a different case, and we will follow up here once we’ve reached a resolution.


The basic library-side setup follows the QuickStart below.

dope_config.yaml also uses the git data as is.
/image is remapped to /rgb_left.

Isaac Sim is ver. 2022.1.1 and uses the robot camera from the carter_warehouse_apriltags_worker.usd.
And the resolution is set to 640x480 in ActionGraph’s ViewPortResolution.
The camera parameters have been adjusted in the parameter in carter_camera_stereo_left to be close to the values of camera_matrix in dope_config.yaml.
The following image shows the setting values.

Thanks for the additional information. We’ve been able to reproduce the DOPE issue and are working on it now.