ArcFace: Additive Angular Margin Loss for Deep Face Recognition
Citations
16 citations
Cites background from "ArcFace: Additive Angular Margin Lo..."
...tion [11] for face image, cos(θj,i) = W j f(xi) ||Wj ||2||f(xi)||2 , and Wj is the weight vector assigned to the i th class....
[...]
...In Equation 4, following ArcFace [11] framework, m1, m2, and m3 are set to 0.9, 0.4, and 0.15, respectively....
[...]
...In Equation 4, following ArcFace [11] framework, m1, m2, and m3 are set to 0....
[...]
16 citations
Cites background from "ArcFace: Additive Angular Margin Lo..."
...Loss Functions Decision Boundaries Softmax [14] (W1 −W2)x+ (b1 − b2) = 0 N-softmax [23] cos θ1 − cos θ2 = 0 CosFace [15] cos θ1 − cos θ2 −m = 0 ArcFace [16] cos(θ1 +m)− cos θ2 = 0 CVM-softmax cos θ1 − cos θ2 − CVM(θ1, θ2) = 0 CVM(θ1, θ2) = m1 sin 2 θ1 +m2 cos2 θ2...
[...]
...Typical margin-based losses are L-softmax [8], A-softmax [14], CosFace [15] and ArcFace [16], which add margins to the angular space or cosine space, respectively....
[...]
...To improve the accuracy of face verification and face identification, CosFace and ArcFace apply a constant margin to cosine and angular spaces, respectively, to make the feature more discriminative....
[...]
16 citations
16 citations
Cites background or methods from "ArcFace: Additive Angular Margin Lo..."
...The face images are detected and aligned by RetinaFace [5], and the cosine similarity matrix is predicted by ArcFace [4]....
[...]
...ArcFace [4] is an additive angular margin loss designed on the softmax loss....
[...]
...Since the margin penalty is a confirmed effective method for deep face recognition and the image-to-image comparison in Triple loss [22] is too tricky, recent improvement works [17, 27, 26, 4] focused on incorporating margin penalty into a more feasible framework, softmax loss, which has extensive imageto-class comparisons within the multiplication step between the embedding feature and the linear transformation matrix....
[...]
...Motivations of translating Triplet loss [22] into Arcface loss [4]....
[...]
...In this report, we first set up the baseline by using our RetinaFace [5] and ArcFace [4]....
[...]
16 citations
References
123,388 citations
33,597 citations
30,843 citations
13,268 citations
10,447 citations