Face Verification.

MOTIVATION:

Face Verification is a problem whereby we are required to confirm if a pair of images depict the same peron’s facial features. This task is widely used in modern day applications like the popular ‘Face-unlock’ feature in smartphones, document id verification etc. This task can essentially be split into two steps, face classification followed by face verification. ResNets have proven to be the state of the art choice for such tasks. Face verification task is accomplished by comparing the embeddings of the pair of images and calculating the similarity of these embeddings. These embeddings are obtained from a trained face classification network.

HOW IS THE VERIFICATION PERFORMED?

As explained above, the entire verification task can be split into two problems.

1. The classification problem, where we will train our neural network to correctly match the image to its corresponding label.
2. The verification problem, where we will verify the cosine similarity between the embeddings of the images passed through the model. These embeddings are the features extracted from the penultimate layer of the model.

Thus the above two problems are tackled by building a ResNet architecture.

WHY RESNET?

ResNets have been known to outperform classical CNNs by allowing us to build deeper networks. The primary shortcomings of classical CNN network like for example VGG-19 network exhibit higher training and validation error as compared to a shallower network. The paper on ResNet https://arxiv.org/pdf/1512.03385.pdf says that “not all systems are similarly easy to optimize.”. This could be because of the problem of vanishing gradients, and this is where ResNets have an advantage. ResNets enable the gradients to reach every layer and thus allows us to build deeper networks with better performance.

EVALUATION:

Once the classification network is trained, the image embeddings are obtained from the penultimate layer of this network. These embeddings are then obtained for a given pair of images from which the Cosine similarity is calculated. This value is obtained for the entire testing data and the ROC-AUC score for this is calculated. The current architecture accomplishes the face verification task with an ROC-AUC score of 93.45.