Matlab codes for evaluation on MegaFace

Requirement:

1. MatMTCNN. If you are not using Windows, you need to modify some of the codes to directly use the Matlab version of MTCNN. If you have done such a work, I'm glad to merge your codes.

2. Megaface. Please download Our dataset, FaceScrub full tgz, FaceScrub bounding boxes actors txt, FaceScrub bounding boxes actresses txt, Linux Development Kit.

Procedure:

1. Align Facescrub: align_facescrub.m. Then mannully confirm some failed samples through align_facescrub_failures.m.

2. Align Megaface: align_megaface_from_list.m. Then align failed samples through align_megaface_failures.m.

3. Extract Features using extract_facescrub_feature.m, extract_megaface_feature.m.

Alignment Logic:

The provided 3-point labels are the most accurate. The second accurate information is the bounding box. All 5-point labels are totally wrong. So our logic is as follows.

1. align_megaface_from_list.m: If there is 3-point label, rotate and crop the image according to it. Then detect and align face from the cropped image.
2. align_megaface_failures.m: If there is no 3-point label or failed to detect from the cropped image, detect face from the raw image.

        If the detected face and the given bounding box's IoU is over 30%, align this face.

        If the IoU is below 30%, use the last two networks of MTCNN to forcely get the face score and 5 keypoints from the cropped image based on the given bounding box.

        If the face score is above 0.3, use the detected 5 points to align the face.

        If all methods are failed, directly crop the middle area of a face as the aligned face.