ArcFace: Additive Angular Margin Loss for Deep Face Recognition
Citations
66 citations
Cites methods from "ArcFace: Additive Angular Margin Lo..."
...Inspired by the studies on metric learning for face recognition [9, 30, 48] that perform metric learning on features on a hypersphere, we normalize the outputs of prediction networks before computing similarities and use a constant value s = 64 [9] to re-scale the features....
[...]
66 citations
Cites background from "ArcFace: Additive Angular Margin Lo..."
...Together, these two steps are the cornerstones of many face-based reasoning tasks, most notably recognition [18, 47, 48, 49, 74, 76] and 3D reconstruction [20, 30, 71, 72]....
[...]
66 citations
Additional excerpts
...3 + R100 [2] + EPolyFace [20] + IncRes-v2 [29] + SE154 [12]...
[...]
66 citations
Cites background from "ArcFace: Additive Angular Margin Lo..."
...[103] reformulated the cost function for face recognition....
[...]
65 citations
Cites background or methods from "ArcFace: Additive Angular Margin Lo..."
...By observing that the weights from the last fully connected layer of a classification DCNN trained on the softmax loss bear conceptual similarities with the centers of each class, the works in [4, 18, 33] proposed several margin losses to improve the discriminative power of the trained model....
[...]
...The two margin losses are: 1) Additive angular margin loss [4], which add an additive angular margin to the angle between the weight vector and feature embeddings....
[...]
...That is, our method involves semantic similarity among classes in meta-training task to learn a more suitable margin penalty, compared with a fixed one generated by [4, 33]....
[...]
...[4] proposed an additive angular margin loss to further improve the discriminative power of feature embedding space....
[...]
References
123,388 citations
33,597 citations
30,843 citations
13,268 citations
10,447 citations