The best face detection algorithm is the Haar-liked (Viola-Jones).

It is a bit slower than the LBP method.

Hameed R. Farhan Thanks for your answer. what about the comparison between Viola-Jones and MTCNN? which is better for using in wild environment?

Viola-jones is very old and computationally very expensive. We use a deep learning based face detection followed by some constraints for movies data.

Follow the link.

https://www.researchgate.net/publication/330234606_DeepStar_Detecting_Starring_Characters_in_Movies?_sg=vbmfJT4FTHe4CV3fKDRnFgQq17NWGHnuxoHNT9KLzCAC_q7DkFc6a29eJqyX2XN_vGQXyHCXkGOkIS5vh95wg5KeZg1j-Puuu01aSHaD.e9q-vGTiT0tWQMSo8BkHVln-wBCfZJxBTvTKX08QI7cRajUIEaoVQfbIsVXQ1iBBocbdQE9YpohbSK_84I-qDw

Really, I didn't examine the MTCNN method.

if you go for performance wise on a lower end Hardware then choose ASM, other wise R-CNN is best if you have high end hardware

In my experience, models got by deep learning are much better than HOG detector of Viola Jones (Haar wavelets) detector. See the first phase of program which can be downloaded at http://www.agentspace.org/andy/learnopencv/

Thanks all. So, with the suggestions given, is it better to use R-CNN or Fast R-CNN? (given that I do not want the test time to be too much, About 3 or 4 frames per second)

The following two survey papers analyzes most of the face detection algorithms in unconstrained environment:

1.Article A Survey of Unconstrained Face Recognition Algorithm and Its...

2.Article Unconstrained Facial Recognition Systems: A Review