the align function may Shear the face image about facenet

[davidsandberg]

Yes, I think this is a known problem which has been discussed for example here. To my knowledge it has not been really solved though.
Have you tried this approach? Does it produce better results?

Another approach would be to not do any transformation at all but instead just use a bounding box and let the CNN handle any rotations etc within that box. Rotations can then be seen as a kind of data augumentation instead. I have tried this approach but using MTCNN for face alignment, and when training a Inception-Resnet-v1 network on this data I can get a model with accuracy ~0.975 on LFW. Not sure how much of the performance improvement that can be attributed to not having the shearing effect though.

[wowo200]

I am not very sure that it can produce better results.
But I use the code :

        assert imgDim is not None
        assert rgbImg is not None
        assert landmarkIndices is not None

        if bb is None:
            bb = self.getLargestFaceBoundingBox(rgbImg, skipMulti)
            if bb is None:
                return

        if landmarks is None:
            landmarks = self.findLandmarks(rgbImg, bb)

        npLandmarks = np.float32(landmarks)
        tplLandmarks = imgDim * MINMAX_TEMPLATE*scale + imgDim*(1-scale)/2
        tplLandmarks = np.transpose(tplLandmarks)
        npLandmarks = np.vstack( (np.transpose(npLandmarks), np.ones(tplLandmarks.shape[1])) )
        #npLandmarkIndices = np.array(landmarkIndices)

        #pylint: disable=maybe-no-member
        #H = cv2.getAffineTransform(npLandmarks[npLandmarkIndices],
        #                           imgDim * MINMAX_TEMPLATE[npLandmarkIndices]*scale + imgDim*(1-scale)/2)
        H = np.matmul(np.matmul(tplLandmarks, np.transpose(npLandmarks)), 
            np.linalg.inv(np.matmul(npLandmarks,np.transpose(npLandmarks))))
        thumbnail = cv2.warpAffine(rgbImg, H, (imgDim, imgDim))

        return thumbnail

this method can get a global transformation matrix and I did not get a better result yet.

[louielu1027]

@davidsandberg Can you open the code of Inception-Resnet-v1 and the model with accuracy ~0.975 on LFW? ~Thank you very much!

[davidsandberg]

All the code for training is already in the repo. I ran the command
python facenet_train_classifier.py --logs_base_dir /media/david/BigDrive/DeepLearning/logs/facenet/ --models_base_dir /media/david/BigDrive/DeepLearning/models/facenet/ --data_dir ~/datasets/facescrub/facescrub_mtcnnalign_182_160:~/datasets/casia/casia_maxpy_mtcnnalign_182_160 --image_size 160 --model_def models.inception_resnet_v1 --lfw_dir ~/datasets/lfw/lfw_mtcnnalign_160 --weight_decay 2e-4 --optimizer RMSPROP --learning_rate -1 --max_nrof_epochs 80 --keep_probability 0.8 --random_crop --random_flip --learning_rate_schedule_file ../data/learning_rate_schedule_classifier_long.txt

With the learning rate schedule (../data/learning_rate_schedule_classifier_long.txt)

# Learning rate schedule
# Maps an epoch number to a learning rate
0:  0.1
65: 0.01
77: 0.001
1000: 0.0001

The Inception-Resnet-v1 will improve the performance significantly compared to the nn4 model but to get to 0.975 a better alignment (MTCNN) is also needed. The code that I used to get the above result can be found here but it requires Caffe installed and to clone the MTCNN repo, which i don't plan to describe here. Instead I'm working on an implementation of this using python/tensorflow but this in not ready yet.

[scotthong]

@davidsandberg

I was studying Dlib recently and found that Dlib already provides the ability to detect faces, extract landmarks, and with one additional step, save the "face_chips" (aligned using the landmark) as image files.

http://dlib.net/imaging.html

I tried this approach to process the LFW dataset and it seems to be working pretty well.
The only caveat is that, the Dlib face detector is not able to detect all the faces in the LFW dataset. For these images, images was cropped directly without alignment using landmark features.

Since "align_dlib.py" is already using Dlib and it seems that the easiest way for this task is to use native Dlib all the way.

There is an example in the Dlib source tree:

http://dlib.net/face_landmark_detection_ex.cpp.html

The following is an example on how to save the aligned face_chips to files.

    for (unsigned long j = 0; j < face_chips.size(); ++j) {
        std::ostringstream stringStream;
        stringStream << "build/tmp/fl_chip_" << j << ".jpg";
        std::string filename = stringStream.str();
        save_jpeg(face_chips[j], filename, 100);
    }

Thanks,

--Scott

[melgor]

@davidsandberg
I see that you upload the code using MTCNN, great!
I have some question about it:

1. Why you do not use detected points to 2D transform the faces, are there any reason?
2. What about the speed of your implementation on TF? I have measured it as about ~25 FPS on VGA image (on the second run of detection, because fist detection is very slow). The author claim to have ~100 FPS using MatLab implementation (do you achive such speed using MatLab?).

Hint: I have replace your "imResample" by "scipy.misc.imresize" and get >50% faster evaluation.

[davidsandberg]

@melgor
Happy to see that you are looking at the MTCNN implementation.
The implementation still contains a couple of bugs which I hope to have solved after the coming weekend, but until then you should use it with care.

1. The main reason is that I'm lazy ;-) and the quickest/easiest way was to just use the bounding boxes. But also, since the MTCNN is better at detecting profile faces I figured that this could cause more severe distortions to the images. But this is an interesting point for investigation that I haven't done yet.
2. I haven't looked at the speed yet. When the debugging of the code is done it would be interesting to check. The impression I got (totally unscientifically) is that it was approximately the same speed as the matlab implementation. But I will have to get back to you on this one...

For the resize thing I guess it looks quite crazy :-), but the reason for using the home-brewed implementation was that while comparing the tensorflow implementation to the matlab one I needed a resample that worked identically in the two implementations. So I ended up having that same code in matlab as well, which should be exchanged for a scipy or opencv implementation when the two implementations match.

[kaishijeng]

I tried with the following command:

python facenet_train_classifier.py --logs_base_dir logs/facenet/ --models_base_dir models/facenet/ --data_dir ./align/casia --image_size 182 --model_def models.inception_resnet_v1 --lfw_dir ./align/datasets/lfw_160 --weight_decay 2e-4 --optimizer RMSPROP --learning_rate -1 --max_nrof_epochs 80 --keep_probability 0.8 --random_crop --random_flip --learning_rate_schedule_file ../data/learning_rate_schedule_classifier_long.txt

It has the following error:

Traceback (most recent call last):
File "facenet_train_classifier.py", line 309, in
main(parse_arguments(sys.argv[1:]))
File "facenet_train_classifier.py", line 100, in main
phase_train=phase_train_placeholder, weight_decay=args.weight_decay)
File "/mnt/2TB/src/facenet/src/models/inception_resnet_v1.py", line 145, in inference
dropout_keep_prob=keep_probability, reuse=reuse)
File "/mnt/2TB/src/facenet/src/models/inception_resnet_v1.py", line 167, in inception_resnet_v1
with tf.variable_scope(scope, 'InceptionResnetV1', [inputs], reuse=reuse):
File "/usr/lib/python2.7/contextlib.py", line 84, in helper
return GeneratorContextManager(func(_args, *_kwds))
TypeError: variable_scope() got multiple values for keyword argument 'reuse'

Do you know what went wrong with my command?
My tensorflow version is 10.0.

Thanks

[davidsandberg]

I'm pretty sure I ran into the same problem but now I'm not sure how it was solved. Could have been that it was fixed when upgrading to a newer version of slim, but I'm not sure.
One option would be to upgrade to TF 0.11 and see if that fixes the problem. I will do that as well as soon as I'm done with some training.
If the problem persists, please file a new issue specifically for this.