Unable to obtain experimental results in the paper（supervised with CR）

[Scanood]

Thank you for your work. I have used your code for training on RTX 3090, and the parameters have not changed. However, I cannot obtain experimental results similar to those in the paper(hand), even if I have trained more epochs. here is my results: If you could provide some details of the training, I would greatly appreciate it.

[stevenlsw]

Hi, are you training on the HO3D-v2 dataset. Could you also check the loss of hand and object to see if they are balanced.

[Scanood]

Hi, are you training on the HO3D-v2 dataset. Could you also check the loss of hand and object to see if they are balanced.

Thank you for your reply. In my experiment, the dataset I used was HO3D_ V2, where the results of the object are better than those in the paper, but the results of the hand posture are not as good as those in the paper. Below are my experimental results.

(epoch: 60, time: Thu Jul 13 12:43:30 2023),mano_mesh3d_loss:0.2526485025882721,mano_joints3d_loss:0.2655828893184662,manopose_loss:0.09159690141677856,manoshape_loss:0.050837427377700806,mano_total_loss:0.6606662273406982,hm_joints2d_loss:0.3534584939479828,obj_reg_loss:6.878592014312744,obj_conf_loss:2.740598440170288,total_loss:10.63332748413086

(epoch: 80, time: Fri Jul 14 01:14:28 2023),mano_mesh3d_loss:0.20092356204986572,mano_joints3d_loss:0.2108018398284912,manopose_loss:0.08086392283439636,manoshape_loss:0.04342929273843765,mano_total_loss:0.5360180139541626,hm_joints2d_loss:0.3074028193950653,obj_reg_loss:6.103377342224121,obj_conf_loss:2.5656049251556396,total_loss:9.512419700622559

(epoch: 100, time: Fri Jul 14 13:14:41 2023),mano_mesh3d_loss:0.17813703417778015,mano_joints3d_loss:0.18687306344509125,manopose_loss:0.07544208317995071,manoshape_loss:0.040135279297828674,mano_total_loss:0.48058727383613586,hm_joints2d_loss:0.2837403118610382,obj_reg_loss:5.711888313293457,obj_conf_loss:2.478605031967163,total_loss:8.95482063293457

I would greatly appreciate it if you could provide some suggestions, which would be very useful to me. Thank you.

[stevenlsw]

Hi, you could try to balance or re-weight the loss of hand and object. Given the two loss are competitive, in supervised learning phase this might happen if one branch is better than the other. And this also motivate us to do semi-supervised learning to stable the performance.

[223d]

Hi, are you training on the HO3D-v2 dataset. Could you also check the loss of hand and object to see if they are balanced.

Thank you for your reply. In my experiment, the dataset I used was HO3D_ V2, where the results of the object are better than those in the paper, but the results of the hand posture are not as good as those in the paper. Below are my experimental results.

(epoch: 60, time: Thu Jul 13 12:43:30 2023),mano_mesh3d_loss:0.2526485025882721,mano_joints3d_loss:0.2655828893184662,manopose_loss:0.09159690141677856,manoshape_loss:0.050837427377700806,mano_total_loss:0.6606662273406982,hm_joints2d_loss:0.3534584939479828,obj_reg_loss:6.878592014312744,obj_conf_loss:2.740598440170288,total_loss:10.63332748413086

(epoch: 80, time: Fri Jul 14 01:14:28 2023),mano_mesh3d_loss:0.20092356204986572,mano_joints3d_loss:0.2108018398284912,manopose_loss:0.08086392283439636,manoshape_loss:0.04342929273843765,mano_total_loss:0.5360180139541626,hm_joints2d_loss:0.3074028193950653,obj_reg_loss:6.103377342224121,obj_conf_loss:2.5656049251556396,total_loss:9.512419700622559

(epoch: 100, time: Fri Jul 14 13:14:41 2023),mano_mesh3d_loss:0.17813703417778015,mano_joints3d_loss:0.18687306344509125,manopose_loss:0.07544208317995071,manoshape_loss:0.040135279297828674,mano_total_loss:0.48058727383613586,hm_joints2d_loss:0.2837403118610382,obj_reg_loss:5.711888313293457,obj_conf_loss:2.478605031967163,total_loss:8.95482063293457

I would greatly appreciate it if you could provide some suggestions, which would be very useful to me. Thank you.

Hi,I also had a similar problem with you，May I ask what problem did you solve to achieve the results in your paper?