Failed segmenetation on side faces and far images
achalshah20 opened this issue · comments
Hi, Thanks for this project. It works good for the front faces. However, for some scenarios results are not so good.
- if person in the image is far from camera(no segmentation at all)
- Side faces (bad segmentation)
Few suggestions or update to this paper will be great help. Thanks!
We treat detection and segmentation as separate problems, the segmentation must follow a detection step. In order to not impose a specific detection implementation, this segmentation assumes an already cropped image around the head.
For best results crop your images using this function (which was also used for training the network):
https://github.com/YuvalNirkin/find_face_landmarks/blob/16cd2443d903518a8dfe417a953851e5ccc47a80/sequence_face_landmarks/utilities.cpp#L291-L344.
For a complete image preprocessing example:
https://github.com/YuvalNirkin/face_swap/blob/64c7eebbb981d87eac851d854e69934451a2f9ff/face_swap/face_swap.cpp#L310-L340
We'll update the documentation for this project to clarify this.
Best,
Yuval
Hi @YuvalNirkin Yes. So I did detection using another model followed by your segmentation model. It works reasonably well with front faces. But it doesn't work with side faces. Is this model trained for side faces as well? Thanks!
The model was trained with side faces as well but naturally their quantity is considerably less.
Could you please upload some examples?
Hi,thanks for your project.i am wondering that if u can provide face segment dataset including photos and labels.its a little difficult to make a train-dataset myself.Thanks~!
@Rettichhuang
Hey, please see #6.
Hi,thanks for your project.i am wondering that if u can provide face segment dataset including photos and labels.its a little difficult to make a train-dataset myself.Thanks~!
Can take a look on this dataset:
http://www.cs.ubbcluj.ro/~dadi/face-hair-segm-database.html