Mirroring and Multi-patch approach about sphereface

[commanderka]

I have to take back my last statement. Concatenating the mirrored features improves the results. Nevertheless, I would be interested in more details concerning the multi patch approach.

[CasiaFan]

@wy1iu Could you tell us a little bit more information about your patch selection? I know one method is to crop original image centered at specific landmarks and train a new net for each.

[wy1iu]

@CasiaFan Yes, you are right. Three-patch ensemble in the paper is to train three networks for three different patches, respectively. In the inference stage, we simply concatenate the features produced by the three networks. We have not used PCA to reduce the dimension, but PCA is optional here (potentially you could get better performance).

[CasiaFan]

@wy1iu Thanks for you response! But how do you choose the patches? Crop patches from aligned face region centered at some particular landmarks then resize to 96x112 before inputting to networks? Or just randomly crop with given size from initial face region?

[wy1iu]

Usually we crop the face images centered at some particular landmarks, since you have to align all your training samples and testing samples. You could use cross-validation to decide how many patches and which patches you want to ensemble.

[JoyLuo]

Hi @wy1iu
"For all experiments, the final representation of a testing face is obtained by concatenating its original face features and its horizontally flipped features." It is to say, the length of 3-patch concatenating feature is 512 * 2 *3 ?
and could you tell me the accuracy position and scale of the patch for reproduce the experiment ?
Thanks a lot.

[JoyLuo]

Hi @CasiaFan
could you tell me the guideline to crop the multi-patch of face ?