CenterPose Inference Shift in Isaac ROS

Robotics & Edge Computing Robotics - Isaac Isaac ROS
1

I have posted this on a GitHub issue as well, but I would like to also post it on this forum.
CenterPose Inference Shift in Isaac ROS · Issue #68 · NVIDIA-ISAAC-ROS/isaac_ros_pose_estimation

In Isaac ROS, there is a horizontal shift in the CenterPose inference coordinates.
When checking the inference results in TAO, no such shift occurs.
isaac ros

image
image450×829 36.8 KB

tao5.5

image
image487×649 29.7 KB

I believe the issue is likely on the Isaac ROS side. Is there any solution or improvement to address this?

2

For now, the issue was resolved by sending a 512×512 image topic instead of 600×800.

3

Hi @k_m

Thank you for your post. If I remember well the model is trained with images 512x512, let me double check with the engineering and I keep you posted.

Raffaello

1 Like
4

Hi @k_m,

I confirm that the model trained is for 512x512 images, which is why changing the image size to the center no longer works.

Evaluation Data

The performance of the CenterPose model during inference was evaluated using the test samples from each category in the official dataset release. These frames, originally high-resolution images of 600x800 pixels, were resized to 512x512 pixels before being processed by the CenterPose model.

I add a note also in our Isaac ROS documentation.

Best,
Raffaello