alfredtorres / 3DFacePointCloudNet

Learning Directly from Synthetic Point Clouds for "In-the-wild" 3D Face Recognition
MIT License
61 stars 15 forks source link

Question about the uploaded pretrained checkpoint #1

Open ggjy opened 4 years ago

ggjy commented 4 years ago

Thanks for your good work! I try to use your pretrained model in train_cls.py, however, when I load the checkpoint and get a low Acc during validation, is there something wrong with the .pth file? I get such results: Test set: Average loss: 10.843366970735438, Accuracy: (0.001176470541395247)

alfredtorres commented 4 years ago

Could you tell me you validate the model on which dataset?

HeatherWyn commented 4 years ago

I try to use your pretrained model in train_triplet to test NN0 vs NNx in the Bophorus Database without fine tune. I deal with the Bophorus Database like this: the viewpoint in PCL is (0, 0, 0), then I change the generated normals to positive z-axis. I use your pretrained model on NN0 vs NNx, and I get such result: top1 = 40.72%, auc = 83.04%, tpr = 0.0464, fpr = 0.0010. In contrast, Without loading the pretrained model, the result is: top1 = 31.96%, auc = 74.64%, tpr = 0.1443, fpr = 0.0014.

In addition, I try to use your pretrained model in train_cls to test the database generated by GPMM with num_class=1000. The diiference is that I didn't change the generated normals to positive z-axis for this database. I get the result like ggjy says.

I think changing the generated normals to positive z-axis is not the key. Could you tell me what's wrong with me?

Thank you very much!