Non-converging region loss

[jabarragann]

Hi,

First of all, thanks for publicly releasing this great work! For my work, I tried to use GDRNet to do 6D pose detection of surgical needles. I adapted my dataset to the BOP format and got the network to train. The trained network is already giving me some significant good results, and the only problem that I noticed is that one of the loss functions for the network (loss_region) never actually improved, which is making the total loss also look bad.

I compared this against a training plot from TUDL, which definitely shows some better behavior. See below the TUDL training plots and my custom dataset training plots. Would anybody be able to give some pointers on why the loss_region looks so bad for my custom dataset? Again, I am already getting some relatively good pose results, but I would definitely want to get the network as optimal as possible.

Some additional details that might or might be relevant:

• The object I am trying to detect is very thin.
• My custom dataset follows very closely the BOP format.

Any help or suggestion would be much appreciated!

My custom dataset training plots

TUDL dataset training plots

[shanice-l]

We generate regions from the object model. You can check the code in data_loader for region generation and visualize them in batch_data_utils.py.

[cnalty]

I'm having a similar problem and I've tried searching for batch_data_utils.py, but I can' see to find a file in the repository with that name

[utsavrai]

I am facing similar problem and could not make it work