Learning Deep Face Representation [18]

Original Abstract

Face representation is a crucial step of face recognition systems. An optimal face representation should be discriminative, robust, compact, and very easy-to-implement. While numerous hand-crafted and learning-based representations have been proposed, considerable room for improvement is still present. In this paper, we present a very easy-to-implement deep learning framework for face representation. Our method bases on a new structure of deep network (called Pyramid CNN). The proposed Pyramid CNN adopts a greedy-filter-and-down-sample operation, which enables the training procedure to be very fast and computation-efficient. In addition, the structure of Pyramid CNN can naturally incorporate feature sharing across multi-scale face representations, increasing the discriminative ability of resulting representation. Our basic network is capable of achieving high recognition accuracy (85.8

Main points

• New deep structure Pyramid CNN
• Labeled Faces in the Wild (LFW)
  • faces
  • 1680 of the people have two or more distinct photos
  • Detected by Viola-Jones detector
  • http://vis-www.cs.umass.edu/lfw/
• State-of-the-art performance on LFW benchmark ()

• Good face representation
  • Identity-preserving: Same person pictures close in feature space
  • Abstract and Compact: from high to low dimensionality
  • Uniform and Automatic: NO hand-crafted and hard-wired parts

• Pyramid CNN
  • ID-preserving Representation Learning: Loss functions measures distance in output feature space
  • Convolutions and Down-sampling
  • Deeper give best results, but increases rapidly the training time
  • Each CNN own private output layer and gets the input from the previous shared layer
  • Only the output of the last level network is used for the represetnation
  • The rest of the outputs is just for training
• Results
  • 164 incorrect predictions
  • Some of them are incorrectly labeled
  • Others are very difficult for humans, because of the age or pose
  • On LFW benchmark achieves state-of-the-art and close to human on croped images
• With ROC curve as a mesure there is an improvement of 0.07-0.12 with Baseline
• Face recognition does not contemplate affine transformations or perspectives,
• Can be difficult to apply in task such as ImageNet, where the object can be in any place and position