Should we apply face alignment for evaluation data?

wujiyang / Face_Pytorch

face recognition algorithms in pytorch framework, including arcface, cosface, sphereface and so on

Apache License 2.0

809 stars 156 forks source link

Should we apply face alignment for evaluation data? #6

Open lymanblue opened 5 years ago

lymanblue commented 5 years ago

Hi~

Thank you for your great work.

Does the reported accuracy result on validation data (e.g. LFW, MegaFace) apply face alignment process (e.g., MTCNN)?

Thank you.

wujiyang commented 5 years ago

@lymanblue All the training and validation/test data have been aligned by MTCNN to the size of 112*112

lymanblue commented 5 years ago

Thanks.

And there is no file like ms1m.py in dataset directory. Would you provide the file in the future? Or is the data loader of ms1m the same with other dataset loader?

Does the training results of the cleaned-MS1M is the same with the MS1M-V2 provided from InsightFace?

Thank you.

wujiyang commented 5 years ago

@lymanblue The cleaned-MS1M I used is provided by DeepGlint, it only has 3.9M images, while InsightFace has a 5.8M cleaned version.
The dataloader of MS1M is as same as the CASIA-WebFace.

lymanblue commented 5 years ago

Therefore, we have to preprocess (e.g., face alignment) the cleaned-MS1M of DeepGlint by ourself with the aid of the msra_lmk file.

On the other hand, if we use the MS1M-V2 (already aligned?) from InsightFace. Can we use the CASIA-WebFace loader directly for training?

Thank you.

lymanblue commented 5 years ago

Is the MS1M-IBUG from InsightFace the cropped and aligned result of the cleaned-MS1M?

wujiyang commented 5 years ago

You can use your own data (MS1M-V2) to train the models directly.

lymanblue commented 5 years ago

For training the model directly from MS1M-V2 from InsightFace.

Do you mean the following steps? (e.g., LFW for validation)

set --train_root to the faces_emore/train.rec in train.py
set --train_file_list to the faces_emore/train.idx in train.py
set --lfw_test to the faces_emore/lfw.bin in train.py
set --lfw_file_list to the path of txt file (i.e., from http://vis-www.cs.umass.edu/lfw/pairs.txt)

Thank you.

wujiyang commented 5 years ago

No, train.rec and train.idx are the mxnet's data format, the dataloader I provided is suitable for images,just like this: 00000
--- 00000-00001.jpg --- 00000-00002.jpg --- 00000-00003.jpg 00001
--- 00001-00001.jpg ---00001-00002.jpg ---00001-00003.jpg

lymanblue commented 5 years ago

Thank you~!

Could we use the the prepare_data.py from https://github.com/TreB1eN/InsightFace_Pytorch to convert the mxnet's format to the specified format? The data format looks similar. (if identical would be better).

wujiyang commented 5 years ago

Well, you can use it to parse the train.rec file to get original images.

ypw-lbj commented 4 years ago

请问您使用预训练模型了吗？ Have you used the pre training model?