search

nkolot / GraphCMR

Repository for the paper "Convolutional Mesh Regression for Single-Image Human Shape Reconstruction"
BSD 3-Clause "New" or "Revised" License
425 stars 67 forks source link

About test performance #22

Closed zhaoyang10 closed 5 years ago

zhaoyang10 commented 5 years ago

Hello, Thank you for releasing your great work. It helps a lot.

 But I have a little trouble in retraining the model.
 I added human3.6m dataset and set hyperparamters as paper reported, and trained the model from scratch. However, I can only get the Reconstruction Error in about 100 mm.

 I wonder whether there are some special tricks used in training procedure?

Thanks!
zhaoyang10 commented 5 years ago

These are losses and training renders. Meanwhile I tried several different hyperparamters, but it didn't work much. image epoch385_1410_nonparam_rend_cv2 epoch383_1410_nonparam_rend_cv2 epoch384_1410_nonparam_rend_cv2

nkolot commented 5 years ago

We did not use any special tricks. I can see from the plots that the camera scale and ty are the same in all images. I've seen this collapse happen a few times, and when this happens I restarted the training and it worked. Usually you can see if this is the case in the first few thousand iterations. Can you try that?

zhaoyang10 commented 5 years ago

Thanks for replying. 'Restart' means resuming or starting from epoch 1?

nkolot commented 5 years ago

Start training from scratch. Can you try to finetune one of the pretrained models? If something is wrong with the data then the metrics should start to become worse.

zhaoyang10 commented 5 years ago

I use h36m and up3d to train the vertices part ( graph_cnn.gc + graph_cnn.shape ), it seems well.

Then fix the parameters, train the camera part ( graph_cnn.camera ) several times, it always collapses to around (0.85, 0.001, 0.2).

zhaoyang10 commented 5 years ago

Start training from scratch. Can you try to finetune one of the pretrained models? If something is wrong with the data then the metrics should start to become worse.

Thanks for replying, I will try the pretrained model today.