We create a very deep convolutional neural network to extract very high-level features of a face for each person. We provide a database of large-scale images of faces from many sources, e.g. web crowdsourcing, our in-house integrated dataset. The database contains a wide variety of races. The inference time of our model is 115 ms on an Intel corei7 6700k processor. Especially in our solution, we save some images from the client's SDK that are selected based on our own algorithm. This feature makes our matching stronger. In addition, we save the checksum of a video to verify duplicate videos.

Our method has shown promising results in the case of large variations in appearance, e.g. pose, age gaps, skin, glass, make-up and beard.