Is it applicable to add landmark detection in this network

[tenggyut]

According to the paper, Faceboxes seems a good replacement of MTCNN in Face Detection Area. But MTCNN has a built in landmark detection, I wonder it is applicable to change faceboxes into a joint multi task network just like MTCNN?

Also Any idea about filling the performance gap between this implementation and the caffe one?

Thanks

[tirtile]

I was try to combine this model with onet in MTCNN to detect face and landmarks, it works well.

[tenggyut]

How to combine onet with faceboxes？use faceboxes's prediction as onet's input?

[tirtile]

Yes. But, change it to a multi task network and retrain it may be better.

[tenggyut]

But the feature map generated by faceboxes is not reused, so may hurt the runtime efficiency?

Also, did you reproduce the performance described in the original paper?

[tirtile]

Yep. No, I haven't retrained yet.

[TropComplique]

Hi. It is a good idea to use onet with FaceBoxes to detect facial landmarks.

But you could also train a simple keypoint detector by yourself. Here is an example of training a simple and fast (~0.5 ms on GTX 1080) 5-keypoints detector: https://github.com/TropComplique/wing-loss (it is not completely finished yet). It is an implementation of this: https://arxiv.org/abs/1711.06753.

I believe that it will be hard to train FaceBoxes for keypoint prediction using multitask loss. Because we will need a lot of training data:
images with a lot of face bounding boxes + keypoints for each face.
But we only have data like this:
images with only one face and keypoints for it. For example, CelebA dataset. And this: images with a lot of face bounding boxes only. For example, WIDER dataset.

And I believe onet is trained on face crops only. I mean, it sees only close face regions during training.