search

ranahanocka / MeshCNN

Convolutional Neural Network for 3D meshes in PyTorch
MIT License
1.56k stars 314 forks source link

Question about classification accuracy #76

Closed MenghaoGuo closed 4 years ago

MenghaoGuo commented 4 years ago

Hi, @ranahanocka

I really appreciate your work ! We know SHREC16 has 16 examples in the train set and 4 in the test set for each class. So it has a total of 120 examples in the test set. According to paper, meshcnn can get 98.6% accuracy on the test set and we can compute it should get (120 x 98.6%) 118.3 (not an integer) correct results which seems wrong. The reason for this is you adopt data augmentation in test set and get more than one samples for the same example. I want to know whether it is correct for using this data augmentation way in test set.

Thanks !

ranahanocka commented 4 years ago

Hi @MenghaoGuo ,

You are right -- this is a typo in the pre-print of the paper, it should have said 99.167 (119/120). In any case, the code / pre-trained weights from this repo have the correct results.

MenghaoGuo commented 4 years ago

Thanks for your reply !