HSE_FaceRec_tf

Tensorflow/Keras small models for face recognition.

• MobileNet-192 (vgg2_mobilenet_2.h5 in Keras and identical vgg2_mobilenet_2.pb in Tensorflow). Model size 13 MB.
• ResNet-50 (vgg2_resnet.pb). Model size 95 MB.

These models were trained on a training set from VGGFace2 dataset using Softmax loss

Please put the corrct path to the LFW directory in DATASET_PATH (line 26 of facerec_text.py). Out simple testing script supports FaceNet, Inception ResNet v1 and InsightFace (ArcFace) models. Please download them from corresponding repositories

We tested these models as feature extractors in 1-NN (nearest neighbour) method with 50% train/test split of several known facial datasets including:

• (UPDATED) 3739 photos of 596 persons from the intersection of LFW (Labeled Faces in the Wild) and YTF datasets with more than one photo. Face identification accuracy (single training image per class): 92.1% (MobileNet), 97.8% (VGG2 ResNet), 96.6% (FaceNet), 88.9% (InsightFace)
• 9164 photos of 1680 persons from LFW with more than one photo. Face identification accuracy (train/test split 0.5): 94.8% (MobileNet), 98.8% (VGG2 ResNet), 97.7% (FaceNet), 92.6% (InsightFace)
• 5396 still photos of 500 subjects from img folder of IJB-A dataset. Face identification accuracy: 88.7% (MobileNet), 90.1% (VGG2 ResNet)