Using Another Face Detection Algorithm

[ahkarami]

I want to use another face detection algorithm instead of MTCNN. As MTCNN accuracy isn't very good for detecting some faces. So, I have used my own implemented Face Detection Algorithm which it's accuracy is better than MTCNN. However, I have faced a new problem. The problem is that I can't get appropriate results with integration of the new face detection method via your face embedding model. I think the problem related to the face alignment method. As your example (in ./deploy/test.py) you have used the MTCNN face detector and then by using its landmark detection you have aligned the face image and then pass it to the face embedding model (e.g., LResNet50E-IR). Unfortunately, my face detector just detect faces and doesen't have any face landmark localization method. Would you please kindly help me to guide me how I can integrate this new face detection method (without landmark localization) into your face embedding models? It is worth nothing that I can integrate it into the face embedding model (i.e., without face alignment) but the results were bad.

[nttstar]

Face alignment is a must. You may consider to use dlib to extract 68 face landmarks.

[JackyWang-001]

@ahkarami I use my own face detector too. Here is my method: (1) Using the pyseeta aligment to get the five landmarks. https://github.com/TuXiaokang/pyseeta (2) Using the function preprocess to get the norm_face. https://github.com/deepinsight/insightface/blob/18b040f951c70d1788034866086b9115f5ffc4df/src/common/face_preprocess.py (3)Pass face image to the face embedding model.

[xizi]

look foward to the face alignement model release

[ahkarami]

Dear @Jacky3213, Thank you for your complete answer. I will test it.

[Ancho5515]

@Jacky3213 Hi, can you share the accuracy and speed under MTCNN and pyseeta? Is the pyseeta better than MTCNN?