最新のML,CV,NLP関連論文読み会@ABEJA
https://abeja-innovation-meetup.connpass.com/event/57686/
Yandong Wen, Kaipeng Zhang, Zhifeng Li and Yu Qiao. center loss ( A Discriminative Feature Learning Approach for Deep Face
Recognition ), eccv, 2016.
http://ydwen.github.io/papers/WenECCV16.pdf
5. どんなもの?
Title: A Discriminative Feature Learning Approach for Deep Face
Recognition
Authors: Yandong Wen, Kaipeng Zhang, Zhifeng Li and Yu Qiao
Publication: ECCV 2016
URL: http://ydwen.github.io/papers/WenECCV16.pdf
Authors call this method ”center loss”
5
17. 技術や手法のキモはどこ?
center loss
17
softmax loss center loss
λ: ハイパーパラメータ, n: 出力層の次元数, m: ミニバッチバッチサイス,x_{i}: ミ
ニバッチにおけるcnnの出力特徴ベクトル, Wとb: cnnの出力後の線形演算の重
みとバイアス, c_{yi}: mini-batchごとに算出され更新されてきたx_{i}と対応する
クラスyの平均特徴ベクトル
26. MNISTにおける検証実験(with center loss)
center lossによる形成される特徴空間の違い
only softmax loss
center loss 26
validation dataではラベルに対応する
ベクトル群のクラス内分散が大きい
クラスごとに偏るような特徴空間を
形成できている
37. 参考文献
● [Krizhevsky+2012] A. Krizhevsky et al. Imagenet classification with deep
convolutional neural networks. In Proceedings of Neural Information
Processing Systems, pp. 1097-1105, 2012.
● [Chopra+2005] S. Chopra et al. Learning a similarity metric discriminatively,
with application to face verication. In Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition, pp. 539-546, 2005.
● [Hoffer+Ailon 2015] E. Hoffer and N. Ailon. DEEP METRIC LEARNING
USING TRIPLET NETWORK. In Proceedings of International Workshop on
Similarity-Based Pattern Recognition, 2015.
● [Wang+2014] J. Wang et al. Learning ne-grained image similarity with deep
ranking. In Proceedings of IEEE Conference on Computer Vision and
Pattern Recognition,pp. 1386-1393, 2014.
37
38. 参考文献
● [Taigman+2014] Y. Taigman et al. Deepface: Closing the gap to
human-level performance in face verification. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. 2014.
● [Wolf+2011] L. Wolf et al. Face recognition in unconstrained videos with
matched background similarity. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2011.
● [Miller+2015] D. Miller et al. Megaface: A million faces for recognition at
scale. arXiv preprint arXiv:1505.02108, 2015.
38
41. center loss v.s. triplet network on MNIST
center loss
training data=50,000
feature vectors of 10,000
validation image
triplet network
trianing data=100
feature vectors of 59,900
validation images 41