search

qiexing / face-landmark-localization

cnn network predict face landmarks (68 points) and head pose (3d pose, yaw,roll,pitch).
410 stars 184 forks source link

instruction of training #3

Closed kaishijeng closed 8 years ago

kaishijeng commented 8 years ago

Your pre-trained model performance is quite good. Can you provide info of how you train your net?

Thanks

qiexing commented 8 years ago

I prepare data ,then fine-tuning the vgg_s caffe model

kaishijeng commented 8 years ago

Which data set do you use to generate a pre-trained model? Can you share you training script?

Thanks,

On Fri, Jan 22, 2016 at 4:37 AM, kkhe notifications@github.com wrote:

I prepare data ,then fine-tuning the vgg_s caffe model

— Reply to this email directly or view it on GitHub https://github.com/qiexing/face-landmark-localization/issues/3#issuecomment-173909529 .

qiexing commented 8 years ago

I use ibug 300w 68 landmark

kaishijeng commented 8 years ago

I have downloaded ibug 300W 68 dataset. Do you mind sharing your training scripts and solver.protobuf?

Thanks,

On Mon, Jan 25, 2016 at 11:03 PM, kkhe notifications@github.com wrote:

Closed #3 https://github.com/qiexing/face-landmark-localization/issues/3 .

— Reply to this email directly or view it on GitHub https://github.com/qiexing/face-landmark-localization/issues/3#event-526078260 .

kaishijeng commented 8 years ago

QieXing

Are you willing to share train.prototxt and solver.prototxt for training purpose?

Thanks

affromero commented 8 years ago

@kaishijeng did you have success reproducing @qiexing results? I've tried to without success. I performed translations and rotations over the 135 original files, as it was suggested, but the model overfit.

I've done this here in case you want to take a look and try yourself in case you have not.

kaishijeng commented 8 years ago

@Afromero

I tried, but not able to make it work. I will take a look at what you have done.

Thanks,

On Mon, Feb 29, 2016 at 8:09 AM, Afromero notifications@github.com wrote:

@kaishijeng https://github.com/kaishijeng did you have success reproducing @qiexing https://github.com/qiexing results? I've tried to without success. I performed translations and rotations over the 135 original files, as it was suggested, but the model overfit.

I've done this here https://github.com/affromero/ActionUnits_Classification/blob/master/Landmarks/face-landmark-localization/get_data.py in case you want to take a look and try yourself in case you have not.

— Reply to this email directly or view it on GitHub https://github.com/qiexing/face-landmark-localization/issues/3#issuecomment-190271545 .

kaishijeng commented 8 years ago

@Afromero,

Did you have a luck of getting  training worked? Your github repository

mentions it is still overfitting.

On Mon, Feb 29, 2016 at 5:14 PM, kaishi Jeng kaishi.jeng@gmail.com wrote:

@Afromero

I tried, but not able to make it work. I will take a look at what you have done.

Thanks,

On Mon, Feb 29, 2016 at 8:09 AM, Afromero notifications@github.com wrote:

@kaishijeng https://github.com/kaishijeng did you have success reproducing @qiexing https://github.com/qiexing results? I've tried to without success. I performed translations and rotations over the 135 original files, as it was suggested, but the model overfit.

I've done this here https://github.com/affromero/ActionUnits_Classification/blob/master/Landmarks/face-landmark-localization/get_data.py in case you want to take a look and try yourself in case you have not.

— Reply to this email directly or view it on GitHub https://github.com/qiexing/face-landmark-localization/issues/3#issuecomment-190271545 .

affromero commented 8 years ago

No @kaishijeng, I have not. I even improved from 135 images to 1M images performing rotations from -80 to 80 degrees and translations and now it does not learn at all, I'm playing with the parameters by now to see what happens.

I hope @qiexing could give us a hand.

kaishijeng commented 8 years ago

@Afromero

Thanks for the update. Let us hope qlexing reads this and shares his training code.

On Mon, Mar 7, 2016 at 6:13 AM, affromero notifications@github.com wrote:

No @kaishijeng https://github.com/kaishijeng, I have not. I even improved from 135 images to 1M images performing rotations from -80 to 80 degrees and translations and now it does not learn at all, I'm playing with the parameters by now to see what happens.

I hope @qiexing https://github.com/qiexing could give us a hand.

— Reply to this email directly or view it on GitHub https://github.com/qiexing/face-landmark-localization/issues/3#issuecomment-193262582 .