通过分析Fashion-Mnist数据集算法基线中的Loss函数,发现问题根因、寻找解决办法、并快速进行算法验证
在基线算法中卷积神经网络,采用了SoftMax函数,
准确率如下:
混淆矩阵如下:卷积神经网络作为基线算法,准确率最高为94.15%。
[1]是CVPR2017的文章,用改进的softmax做人脸识别,改进点是提出了angular softmax loss(A-softmax loss)用来改进原来的softmax loss。 M=2、4、6 的准确率如下:
M=2 混淆矩阵如下:
SphereFace在M=2时,准确率最高为95.32%,比卷积神经网络基线版本,提升与1.17%,并且解决了部分分类不准确的问题。
论文[2]原名是ArcFace,但是由于与虹软重名,后改名为Insight Face,截止2018年3月,是MegaFace榜第一,达到了98.36%的成绩。 M=0.5、2、4 的准确率如下:
M=0.5 混淆矩阵如下:
ArcFace在M=0.5时,准确率最高为94.28%,比卷积神经网络基线版本,提升与0.13%。- 4uiiurz1/keras-arcface
- auroua/InsightFace_TF
- YunYang1994/SphereFace
- SphereFace论文学习
- wujiyang/Face_Pytorch
- Kakoedlinnoeslovo/center_loss
- clcarwin/sphereface_pytorch
- YirongMao/softmax_variants
- [1] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, Le Song. "SphereFace: Deep Hypersphere Embedding for Face Recognition"
- [2] Jiankang Deng, Jia Guo, Niannan Xue, Stefanos Zafeiriou. "ArcFace: Additive Angular Margin Loss for Deep Face Recognition"
- [3] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, Wei Liu. "CosFace: Large Margin Cosine Loss for Deep Face Recognition"
- [4] Liu, Weiyang and Lin, Rongmei and Liu, Zhen and Liu, Lixin and Yu, Zhiding and Dai, Bo and Song, Le. "Learning towards Minimum Hyperspherical Energy" (SphereFace+ is described in Section 5.2 of the main paper)
- [5] Zhong, Zhun and Zheng, Liang and Kang, Guoliang and Li, Shaozi and Yang, Yi. "Random Erasing Data Augmentation"