The RAF-DB that aligned by MTCNN.

msy1412 commented 1 year ago

Excuse me, could you provide the RAF-DB that you train and test? Because I can't achieve the performance of 91.98% on the official aligned test set of RAF-DB with the weight APViT_RAF-3eeecf7d.pth. Then I try to re-align the official aligned test set by MTCNN, but it doesn't work either. I only get the top-1 accuracy: 81.42%.

sbelharbi commented 1 year ago

hi, can you please, at least, provide the full code you used for face alignment with mtcnn? it is an important step and it will help reproduce your results.

very much appreciated thanks

youqingxiaozhua commented 1 year ago

Thank you for your suggestion. I have made the following updates on the Readme: specified the alignment codes, and included a copy of our preprocessed version of the RAF-DB. Please refer to the updated (Readme file)[https://github.com/youqingxiaozhua/APViT/blob/main/README.md] for more infromation.

sbelharbi commented 1 year ago

thanks for the update. what do you do when multiple faces are detected or none is detected? i used face detectors on raf-db, sometimes there are many and sometimes it fails to detect any faces. it would be very helpful to use the same script you used. in your updated readme, you just pointed to https://github.com/ZhaoJ9014/face.evoLVe/#Face-Alignment, unless i misread it.

thanks again for your time

youqingxiaozhua commented 1 year ago

For cases where the face is not detected, I use images from the dataset’s original aligned folder. For cases where there are multiple faces, I am not sure whether I used the original image or just choose the most confident one. Sorry about that, but I I cannot remember that. However, currently, I think the prior strategy looks more reasonable.

sbelharbi commented 1 year ago

thanks. this is helpful. so i assume you no longer have these scripts.

thanks again

sbelharbi commented 1 year ago

hi, i run initial test with the code you mentioned in readme to align faces: face.evolve to crop and align some raf-db faces. python face_align.py -source_root './source' -dest_root './destination' -crop_size 224

i used the code as it is. i run it with few samples from raf-db. i am not getting exactly your images you published here.

it seems that face.evolve order faces from more likely to less likely.
following what you said above does not lead to the same images. this is a problem. can you please dig and try to find your same script you used to build your datasets, so we can use the same to build other datasets instead of distributing datasets like you did (not sure we are allowed to do that...). see below. it would be easier to provide the code.

thanks again for the help

test_0001:

test_0001