search

Vegetebird / MHFormer

[CVPR 2022] MHFormer: Multi-Hypothesis Transformer for 3D Human Pose Estimation
MIT License
530 stars 85 forks source link

MPI-INF-3DHP Dataset training; not getting results as reported in paper #101

Closed MehwishG closed 1 year ago

MehwishG commented 1 year ago

Dear Author,

I have trained and evaluated the proposed Mhforrmer on the Human 3.6m dataset and achieved the same MPJPE as reported in the paper. For MPI-INF-3DHPE dataset, I have followed data pre-processing and loading from P-STMO paper and trained on the MPI-INF-3DHP dataset, used train and test .npz file provided by P-STMO paper. I have tried 9,27,81 and 251 frames but no success here are the train log for 81-frames: Loss is decreasing at train time but test MPJPE is high 2023/05/22 19:13:47 epoch: 1, lr: 0.0010000, loss: 83.8849, p1: 369.81 2023/05/22 20:17:52 epoch: 2, lr: 0.0009500, loss: 58.9413, p1: 370.44 2023/05/22 21:22:29 epoch: 3, lr: 0.0009025, loss: 55.9185, p1: 368.76 2023/05/22 22:28:05 epoch: 4, lr: 0.0008574, loss: 54.2971, p1: 369.98 2023/05/22 23:33:52 epoch: 5, lr: 0.0008145, loss: 53.2584, p1: 368.76 2023/05/23 00:40:14 epoch: 6, lr: 0.0004073, loss: 51.7462, p1: 368.26 2023/05/23 01:47:03 epoch: 7, lr: 0.0003869, loss: 51.3817, p1: 370.00 2023/05/23 02:54:03 epoch: 8, lr: 0.0003675, loss: 51.0777, p1: 369.01 2023/05/23 04:01:10 epoch: 9, lr: 0.0003492, loss: 50.8506, p1: 368.69 2023/05/23 05:08:38 epoch: 10, lr: 0.0003317, loss: 50.5766, p1: 368.93 2023/05/23 06:16:23 epoch: 11, lr: 0.0001659, loss: 50.0719, p1: 368.64 2023/05/23 07:24:34 epoch: 12, lr: 0.0001576, loss: 49.9637, p1: 369.53 2023/05/23 08:32:25 epoch: 13, lr: 0.0001497, loss: 49.8830, p1: 368.41 2023/05/23 09:40:32 epoch: 14, lr: 0.0001422, loss: 49.7708, p1: 368.42

Please guide me, i am trying for some time and not able to achieve results mentioned in paper.

yongxin-Cui commented 1 year ago

Dear Author,

I have trained and evaluated the proposed Mhforrmer on the Human 3.6m dataset and achieved the same MPJPE as reported in the paper. For MPI-INF-3DHPE dataset, I have followed data pre-processing and loading from P-STMO paper and trained on the MPI-INF-3DHP dataset, used train and test .npz file provided by P-STMO paper. I have tried 9,27,81 and 251 frames but no success here are the train log for 81-frames: Loss is decreasing at train time but test MPJPE is high 2023/05/22 19:13:47 epoch: 1, lr: 0.0010000, loss: 83.8849, p1: 369.81 2023/05/22 20:17:52 epoch: 2, lr: 0.0009500, loss: 58.9413, p1: 370.44 2023/05/22 21:22:29 epoch: 3, lr: 0.0009025, loss: 55.9185, p1: 368.76 2023/05/22 22:28:05 epoch: 4, lr: 0.0008574, loss: 54.2971, p1: 369.98 2023/05/22 23:33:52 epoch: 5, lr: 0.0008145, loss: 53.2584, p1: 368.76 2023/05/23 00:40:14 epoch: 6, lr: 0.0004073, loss: 51.7462, p1: 368.26 2023/05/23 01:47:03 epoch: 7, lr: 0.0003869, loss: 51.3817, p1: 370.00 2023/05/23 02:54:03 epoch: 8, lr: 0.0003675, loss: 51.0777, p1: 369.01 2023/05/23 04:01:10 epoch: 9, lr: 0.0003492, loss: 50.8506, p1: 368.69 2023/05/23 05:08:38 epoch: 10, lr: 0.0003317, loss: 50.5766, p1: 368.93 2023/05/23 06:16:23 epoch: 11, lr: 0.0001659, loss: 50.0719, p1: 368.64 2023/05/23 07:24:34 epoch: 12, lr: 0.0001576, loss: 49.9637, p1: 369.53 2023/05/23 08:32:25 epoch: 13, lr: 0.0001497, loss: 49.8830, p1: 368.41 2023/05/23 09:40:32 epoch: 14, lr: 0.0001422, loss: 49.7708, p1: 368.42

Please guide me, i am trying for some time and not able to achieve results mentioned in paper.

Hi, I am also want to trian on the MPI-INF-3DHPE dataset, would you like to share your method to me ? I want to have a look on your method changed based on P_STMO.Thanks a lot.

GloryyrolG commented 1 year ago

Dear Author, I have trained and evaluated the proposed Mhforrmer on the Human 3.6m dataset and achieved the same MPJPE as reported in the paper. For MPI-INF-3DHPE dataset, I have followed data pre-processing and loading from P-STMO paper and trained on the MPI-INF-3DHP dataset, used train and test .npz file provided by P-STMO paper. I have tried 9,27,81 and 251 frames but no success here are the train log for 81-frames: Loss is decreasing at train time but test MPJPE is high 2023/05/22 19:13:47 epoch: 1, lr: 0.0010000, loss: 83.8849, p1: 369.81 2023/05/22 20:17:52 epoch: 2, lr: 0.0009500, loss: 58.9413, p1: 370.44 2023/05/22 21:22:29 epoch: 3, lr: 0.0009025, loss: 55.9185, p1: 368.76 2023/05/22 22:28:05 epoch: 4, lr: 0.0008574, loss: 54.2971, p1: 369.98 2023/05/22 23:33:52 epoch: 5, lr: 0.0008145, loss: 53.2584, p1: 368.76 2023/05/23 00:40:14 epoch: 6, lr: 0.0004073, loss: 51.7462, p1: 368.26 2023/05/23 01:47:03 epoch: 7, lr: 0.0003869, loss: 51.3817, p1: 370.00 2023/05/23 02:54:03 epoch: 8, lr: 0.0003675, loss: 51.0777, p1: 369.01 2023/05/23 04:01:10 epoch: 9, lr: 0.0003492, loss: 50.8506, p1: 368.69 2023/05/23 05:08:38 epoch: 10, lr: 0.0003317, loss: 50.5766, p1: 368.93 2023/05/23 06:16:23 epoch: 11, lr: 0.0001659, loss: 50.0719, p1: 368.64 2023/05/23 07:24:34 epoch: 12, lr: 0.0001576, loss: 49.9637, p1: 369.53 2023/05/23 08:32:25 epoch: 13, lr: 0.0001497, loss: 49.8830, p1: 368.41 2023/05/23 09:40:32 epoch: 14, lr: 0.0001422, loss: 49.7708, p1: 368.42 Please guide me, i am trying for some time and not able to achieve results mentioned in paper.

Hi, I am also want to trian on the MPI-INF-3DHPE dataset, would you like to share your method to me ? I want to have a look on your method changed based on P_STMO.Thanks a lot.

Hi @MehwishG , @yongxin-Cui , sorry, but isn't the result of 58.0 in Tab. 3 achieved by the model trained on 3DHP instead of Human3.6M? Thx & regards,

MehwishG commented 1 year ago

@yongxin-Cui , I have trained MH-Former in PSTMO repo and it worked for me. Simply replace PSTMO model with MH-Former

yongxin-Cui commented 1 year ago

@yongxin-Cui , I have trained MH-Former in PSTMO repo and it worked for me. Simply replace PSTMO model with MH-Former

Thank you for your reply, my friend. I would like to ask about the details of MHFRORMER training on the MPI-INF-3DHP dataset.This is my email 2594348199@qq.com, could you give me some advice ? Thank you very much.

MehwishG commented 1 year ago

Sure, I will email you and will be happy to help