Possible bug in calculation of absolute translation coordinates

[phquentin]

Hi,

when you are transforming the relative scale-invariant translation parameter into absolute coordinates in these lines of code:

https://github.com/THU-DA-6D-Pose-Group/GDR-Net/blob/9814a182e7e4d14431088eab2fea6747616b9bc6/core/gdrn_modeling/models/pose_from_pred_centroid_z.py#L175C5-L182C6

it seems that you use the width and height of the initial detected bounding box in the variable roi_whs coming from here:

https://github.com/shanice-l/gdrnpp_bop2022/blob/f3ca18632f4b68c15ab3306119c364a0497282a7/core/gdrn_modeling/datasets/data_loader_online.py#L470C9-L482C49

instead of the width and height of the shifted one, which is the input to the network. Is this intentional?

Best,

Philipp

[wangg12]

Hi, thanks for pointing this out!
The implementation actually follows the prior work CDPN, and I did not ablate using the scales after shift/scale. In my view, using gt box size can encode the true size prior to the delta center learning. Thus if the predicted bbox is closer to gt bbox, the results would be better.
BTW, if you find using the augmented size can lead to better results, please do let me know.

[phquentin]

Thanks for the quick response, I will let you know.