Hi, thank you for sharing your work and providing support. I was able to run FoundationPose on the provided demo data and was able to reproduce the performance presented in the article regarding the YCBInEOAT dataset.

I am now trying to run FoundationPose on a custom dataset. To do so, I am generating simple synthetic scenes in Blender using the YCB models. When running FoundationPose on the generated scenes, I have much lower performance, with an average AUC (ADD) of 73% compared to the 96% obtained in the provided demo data. This decline in performance makes the translation error be around a few cm rather than a few mm. I used the same model as the demo data, so I doubt the model itself is the error. Here is my synthetic scene and my debug folder:

https://drive.google.com/drive/folders/1f9UaLt_BMojAE-Xi5PyJ_GPHzoQRBWoW?usp=sharing

I would greatly appreciate it if you could take the time to check the debug file and see what I have missed that causes the significant decline in performance. Here are things I have checked and that don't seem to be the problem:

• The texture of the object is correctly applied.
• The scale of the object is correct.
• Camera intrinsics are on the correct scale.
• I've tried with different objects.
• I have checked the quality of the depth data; it seems correct, but that is the part I am the least certain about.

I hope you can help me figure out the decline in performance.

Have a good day!