Know You at One Glance: A Compact Vector Representation for Low-Shot Learning

Low-shot face recognition is a very challenging yet important problem in computer vision. The feature representation of the gallery face sample is one key component in this problem. To this end, we propose an Enforced Softmax optimization approach built upon Convolutional Neural Networks (CNNs) to produce an effective and compact vector representation. The learned feature representation is very helpful to overcome the underlying multi-modality variations and remain the primary key features as close to the mean face of the identity as possible in the high-dimensional feature space, thus making the gallery basis more robust under various conditions, and improving the overall performance for low-shot learning. In particular, we sequentially leverage optimal dropout, selective attenuation, ℓ2 normalization, and model-level optimization to enhance the standard Softmax objective function for to produce a more compact vectorized representation for low-shot learning. Comprehensive evaluations on the MNIST, Labeled Faces in the Wild (LFW), and the challenging MS-Celeb-1M Low-Shot Learning Face Recognition benchmark datasets clearly demonstrate the superiority of our proposed method over state-of-the-arts. By further introducing a heuristic voting strategy for robust multi-view combination, and our proposed method has won the Top-1 place in the MS-Celeb-1M Low-Shot Learning Challenge.