search

sfzhang15 / FaceBoxes

FaceBoxes: A CPU Real-time Face Detector with High Accuracy, IJCB, 2017
Apache License 2.0
592 stars 171 forks source link

Facial landmarks detection #27

Closed RunningLeon closed 5 years ago

RunningLeon commented 5 years ago

Hi, Great job. Its cpu real-time speed is really fast. Wonder how to add facial landmarks detection on your net. Do you have any schedule to release this kind of work? Thanks a lot.

sfzhang15 commented 5 years ago

@RunningLeon We use CelebA and WIDER FACE to jointly train the model. Comparing to the detection, we plus a new branch same as the regression branch to do facial landmarks detection, i.e., every anchor has five initial landmarks to regress the gt landmarks. The schedule to release landmark model depends on our lab teachers.

liny23 commented 5 years ago

@sfzhang15 Is the accuracy of the relevant data correct?

sfzhang15 commented 5 years ago

@liny23 I don't understand your question very well, please describe your problem more clearly

liny23 commented 5 years ago

paper about landmark and result in public dataset

sfzhang15 commented 5 years ago

@liny23 The paper about joint face and landmark detection will be public soon.

onzone commented 5 years ago

On which data you are training your landmarks on?? For jointly learning face detection and landmark detection you should have face detection as well as landmark labels...Wider face has face detection labels and other landmark detection dataset is not good for face detection(either only one face is there in the image or if multiple faces are present in the image only landmark of one face is given)..So are there any dataset that is good for joint training??

sfzhang15 commented 5 years ago

@onzone We use WIDER FACE and CelebA to jointly train face and landmark detection. WIDER FACE only has the face bbox so it just compute the classification and regression loss, while CelebA has the face bbox and landmard so it compute all the losses. We use some data augmentation to enrich the faces in CelebA. The details refer to our forthcoming paper.

onzone commented 5 years ago

@sfzhang15 Thank you for your feedback.