Mis-Classified Vector Guided Softmax Loss for Face Recognition

Abstract

Face recognition has witnessed significant progress due to the advances of deep convolutional neural networks (CNNs), the central task of which is how to improve the feature discrimination. To this end, several margin-based (e.g., angular, additive and additive angular margins) softmax loss functions have been proposed to increase the feature margin between different classes. However, despite great achievements have been made, they mainly suffer from three issues: 1) Obviously, they ignore the importance of informative features mining for discriminative learning; 2) They encourage the feature margin only from the ground truth class, without realizing the discriminability from other non-ground truth classes; 3) The feature margin between different classes is set to be same and fixed, which may not adapt the situations very well. To cope with these issues, this paper develops a novel loss function, which adaptively emphasizes the mis-classified feature vectors to guide the discriminative feature learning. Thus we can address all the above issues and achieve more discriminative face features. To the best of our knowledge, this is the first attempt to inherit the advantages of feature margin and feature mining into a unified loss function. Experimental results on several benchmarks have demonstrated the effectiveness of our method over state-of-the-art alternatives. Our code is available at http://www.cbsr.ia.ac.cn/users/xiaobowang/.

Deep Speaker Embedding Extraction with Channel-Wise Feature Responses and Additive Supervision Softmax Loss Function

“…It is intuitive that the mis-classified samples contribute more to the improvement of the identity discriminability [16]. Given this, we proposed the additive supervision softmax loss (ASsoftmax) to make full use of the prior knowledge of the misclassified samples.…”

“…Moreover, the end-to-end loss like TE2E [14] and GE2E [15] have been proposed to training the speaker model in an end-to-end fashion. It is worth noting that aforementioned loss functions did not pour much attention on the hard samples, which are beneficial for learning a distinctive representation [16,17].…”

A joint loss function for deep face recognition

“…. , w K } and the feature x of the last fully connected layer are usually normalized and their magnitudes are replaced as a scale parameter s (Wang et al, 2017;Deng et al, 2019;Wang et al, 2019b). In consequence, given an input feature vector x with its ground truth label y, the original softmax loss Eq.…”

“…where cos(θ w k ,x ) = w T k x is the cosine similarity and θ w k ,x is the angle between w k and x. As pointed out by a great many studies (Liu et al, 2016;Wang et al, 2018b;Deng et al, 2019;Wang et al, 2019b), the learned features with softmax loss are prone to be separable, rather than to be discriminative for face recognition.…”

“…The task of face recognition contains two categories: face identification to classify a given face to a specific identity, and face verification to determine whether a pair of face images are of the same identity. In recent years, the advanced face recognition methods (Simonyan & Andrew, 2014;Guo et al, 2018;Wang et al, 2019b;Deng et al, 2019) are built upon convolutional neural networks (CNNs) and the learned high-level discriminative features are adopted for evaluation. To train CNNs with discriminative features, the loss function plays an important Proceedings of the 37 th International Conference on Machine Learning, Vienna, Austria, PMLR 119, 2020.…”