Pose estimation seems good, but mode_tf.obj is strange

[MarchesiGiorgio]

Hello, Thank you for your amazing work on FoundationPose – it’s been incredibly useful for my research. I'm currently using the system on a custom dataset, and the pose estimation visually looks correct in the rendered debug outputs (as shown in the first figure below). However, when I inspect the model_tf.obj file (second image below) generated during the first registration, I notice that the 3D model is not aligned as expected.

Why is the model positioned and oriented the way it is in model_tf.obj? Is it placed using the estimated object pose? If so, relative to what origin or frame?

I'm asking because when I use this estimated pose (after proper conversion to Unity’s coordinate system), the object appears in the wrong place – matching the model_tf.obj output, which seems to be shifted and rotated incorrectly compared to the actual scene.

I’ve included a link to my dataset and the debug folder here: https://drive.google.com/drive/folders/1461HItukQhbzQ9pRPtIVuGyd_Mpa60Lt?usp=drive_link

Thank you in advance for any clarification you can provide!

[Manu752]

Hi long did it take to train the model to recognise this object?

[MarchesiGiorgio]

Hi, thanks for your answer. Actually, I did not retrain the network. I just ran run_demo.py with my own dataset. If you’re referring to the time it takes to start the localization, it takes around 30 seconds.

[yys-abu]

Hi, thanks for your answer. Actually, I did not retrain the network. I just ran with my own dataset. If you’re referring to the time it takes to start the localization, it takes around 30 seconds.run_demo.py

Hello, your link is no longer valid. Could you please re-upload it? May I ask what modifications you made? I have uploaded the file shown in the image below:

But running run_demo.py doesn't work well: