Our face recognition algorithm has been reached to the top level of 99.83% in LFW benchmark.

LFW database consists of face photographs designed for researching the issues of unconstrained face recognition.

The dataset contains more than 13,000 images of celebrities collected from the web.

Most of other databases used for face recognition benchmark have been collected by controlling parameters which affect the face recognition such as pose, lighting, expression, background, camera quality, occlusion, age, and gender.

However the photographs of the LFW database include all the natural variants that are frequently encountered by people in real life.

In LFW benchmark, many face recognition algorithms from the world’s famous corporations, universities and institutions like Google and Baidu are being introduced making it the most popular face recognition benchmark nowadays.

In LFW benchmark, face recognition algorithms distinguish that the randomly chosen pairs among face images are belonged to the same person or not.

We used training dataset containing 3.1M images of 59K persons collected from the Internet, which has no intersection with LFW dataset.

We trained six deep CNN models, one of which is the Inception-Resent-like network and the other five are Resnet-50 network, each operating on a different face patch.

These models achieved 99.78%, 99.72%, 99.50%, 98.97%, 99.48% and 99.28% accuracy, respectively.

The feature vectors from six deep CNN models are concatenated and transformed to low-dimensional vector by metric learning.

After feature extraction, the distance between two feature vectors is measured by Euclidean distance.

The following table shows performances of top level state-of-the-art face recognition algorithms in LFW benchmark.

Our face recognition algorithm was tested under Unrestricted, Labeled Outside Data protocol of LFW benchmark.

This protocol is to evaluate the accuracy in the 6000-pair verification tasks.

Figure 1 shows examples of correctly classified pairs from LFW database by our face recognition algorithm.

The example images shown in Figure 1 are considered difficult to recognize because of various pose, expression and occlusion.

Figure 2 shows all misclassified pairs.

Our algorithm produced 10 misclassified pairs out of 6,000 and 5 of them are mislabeled in the database, which means the actual number of misclassified pairs is 5.

(b) False Accepted Pairs

Figure 2. Pairs Failed in LFW benchmark. (a) False Reject Pairs. (b) False Accept Pairs. The score under the image is the distance between two feature vectors from above pair.

In the figure, the ones in red boundary are mislabeled pairs.