MTCNN cascaded to Face Alignment Network

[dpnvnt]

Hi everyone, I don't know if this is the right repo to ask this question but I'm new to GitHub...so let's try!

Me and my university project group have developed a small project for face detection and compared different solutions, among which we have analyzed the MTCNN cascaded to the Face Alignment Network (the output of the MTCNN in input to the Face Alignment Network). We analyzed the RAM occupation of the GPU, additionally evaluated the velocity of the network and compared the results to the ones obtained by only using the Face Alignment Network. The result was that the cascade of the two networks occupies less RAM than the use of the Face Alignment Network alone. Moreover, the cascade is faster than the Face Alignment Network used alone.

We have hypothesized that it could be due to the MTCNN that reduces the size of the output image to 160x160x3 and gives it at the input of the Face Alignment Network but we are not sure about it...

Could someone have some ideas?

Thank you in advance, I will appreciate very much all the answers!