search

microsoft / Deep3DFaceReconstruction

Accurate 3D Face Reconstruction with Weakly-Supervised Learning: From Single Image to Image Set (CVPRW 2019)
MIT License
2.18k stars 443 forks source link

Why the train subset and val subset are exactly the same set folder? #138

Closed HulkMaker closed 3 years ago

HulkMaker commented 3 years ago

Why the train subset and val subset are exactly the same set folder? Will it decrease the reconstruction performance?

YuDeng commented 3 years ago

The training set and validation set shown in the code are just examples. In your experiment, you have to set them to your custom folders. In that case the training set and validation set should be different folders.

HulkMaker commented 3 years ago

thanks for answer！ I set train and val subset different, train:val = 5:1, after 200k iterations, the landmark loss is still high as below, what's the issue? image

YuDeng commented 3 years ago

Here the printed landmark loss is in pixel scale so that a value of 2 means 2-pixel distance to the ground truth, which is reasonable.

HulkMaker commented 3 years ago

great answer！