Not getting the same result as offline foundationpose

Hello,

I am trying to see if Isaac ROS Pose Estimation can be used safely for estimating pose of a door lever but seems like it's breaking at some point.

There are some differences between the original offline FoundationPose and the Isaac ROS FoundationPose which I have been trying to investigate.

The original offline FoundationPose does:

• Initial registration with iterative refinement (iteration=5)
• Tracking with iterative refinement (track_one(... iteration=2))

In contrast, the Isaac ROS FoundationPose registration graph also supports iterative refinement (refine_iterations) but that is only for pose registration (estimation) not for tracking.

• It seems to do a single render --> refine --> transform pass per frame
• Found no exposed tracking refine_iterations parameter in the runtime tracking graph
• Tracking output is recursively fed back as the next tracking input

Looks like for thin/small objects like door lever, small residual tracking error --> reused next frame --> error accumulates --> pose drifts away over time.

Initial Analysis:

• Registration/Initial Pose Estimation output looks good
• Tracking output gradually drifts farther away
• Increasing registration refine iterations in Isaac ROS did not improve tracking drift
• Continuous reinitialization because of tracking drift (Set a threshold limit to reinitialize if the tracking drift is too much).

Isaac ROS Pose Estimation Results:
https://www.youtube.com/watch?v=V2CSoLekMIY
https://www.youtube.com/watch?v=5Dz9PhT-LBM

FoundationPose Offline Pose Estimation Results:
https://www.youtube.com/watch?v=SyeEkQUFTsQ
https://www.youtube.com/watch?v=B6aruRUEm5E

The same data has been fed to both of them but the result is different.