Data projecion problem

[DavidJimenezS]

Dear @ybkscht ,

I've generated some data in unity with the flip Y to make it equal to openCV.

However, my rotations and translation when projected into the image does not match the objects. Do you have any idea what is missing from my side or what could be causing the issue?

[madhanuman]

I am right now also trying to work out the right projections on my image... But I could you exlpain your problem a bit more? Did you already do the training? Or are you using the debug.py file? And also it is important to have the right linemod format... You need to have all the necessary parameters inside the info.yml, gt.yml and model_info.yml files If youre unsure about the linemod structure, download the sample linemod_preprocessed dataset and try to reverse engineer... atleast thats what I did and I am now on the point where I can see first (not yet perfectly alligned results with inference.py)

[DavidJimenezS]

@madhanuman Thanks for your reply. Maybe there was a lack of information from my side. So, yeah, I have data structure equal as the sample the authors provide and I'm able to train.

Nonetheless, the problem is when I check my data with debug file the 3D bbox projection to the image does not match. Since my data comes from unity, I already flip Y axis, but still not working.

Have any idea, what's the problem?

[madhanuman]

Ok I see... The debug.py file uses generators/linemod.py (if you're using linemod) to generate the model... Make sure that you have updated the path files inside linemod.py to your dataset. That was a mistake i made... Using "python debug.py --annotations --draw_2d-bboxes --draw-class-names linemod Datasets\linemod_preprocessed --object-id 1" you should see one additional 2d box which should be directly extracted from the model_info.yml... that helps to see where the problem is... I hope it is clear what i want to say