Loss_region unable to converge

[lu-ming-lei]

Other Loss has significant decline， but Loss_region‘s drop is very weak. My training use config : configs/gdrn/lm/a6_cPnP_lm13.py
Region area choose 4, 16, 64 can not make any improve.

[wangg12]

How is your test performance?

[lu-ming-lei]

The performanceis similar to your result

[lu-ming-lei]

As shown in Paper， without using Loss_region can still work well, but I wonder why it is hard to estimate this part. I think estimate crop_xyz is a harder task and it can work.

[wangg12]

As shown in Paper， without using Loss_region can still work well, but I wonder why it is hard to estimate this part. I think estimate crop_xyz is a harder task and it can work.

Loss region can be regarded as a regularizer or auxiliary task for learning the symmetries and the coarse region information. To estimate a 6D pose, we only need the correspondences, so the network might pay more attention to optimize towards the xyz part rather than the regions. Besides, when doing the class-agnostic training, the learning of regions might be much harder.

[lu-ming-lei]

thanks for your explanation