CosFace: Large Margin Cosine Loss for Deep Face Recognition

CosFace: Large Margin Cosine Loss for Deep Face Recognition

3 Apr 2018 | Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li*, and Wei Liu*
The paper "CosFace: Large Margin Cosine Loss for Deep Face Recognition" by Hao Wang et al. introduces a novel loss function called Large Margin Cosine Loss (LMCL) to enhance the discriminative power of deep features in face recognition tasks. The authors reformulate the traditional softmax loss as a cosine loss by normalizing both feature and weight vectors to remove radial variations. This normalization step, combined with a cosine margin term, maximizes the decision margin in the angular space, achieving minimum intra-class variance and maximum inter-class variance. The proposed model, referred to as CosFace, is evaluated on popular face recognition datasets such as LFW, YTF, and MegaFace, demonstrating state-of-the-art performance. The paper also provides theoretical analysis and comparisons with other loss functions, highlighting the effectiveness of LMCL in improving the discriminative power of learned features.The paper "CosFace: Large Margin Cosine Loss for Deep Face Recognition" by Hao Wang et al. introduces a novel loss function called Large Margin Cosine Loss (LMCL) to enhance the discriminative power of deep features in face recognition tasks. The authors reformulate the traditional softmax loss as a cosine loss by normalizing both feature and weight vectors to remove radial variations. This normalization step, combined with a cosine margin term, maximizes the decision margin in the angular space, achieving minimum intra-class variance and maximum inter-class variance. The proposed model, referred to as CosFace, is evaluated on popular face recognition datasets such as LFW, YTF, and MegaFace, demonstrating state-of-the-art performance. The paper also provides theoretical analysis and comparisons with other loss functions, highlighting the effectiveness of LMCL in improving the discriminative power of learned features.
Reach us at info@study.space