ArcFace: Additive Angular Margin Loss for Deep Face Recognition
Citations
2 citations
2 citations
2 citations
Cites background or methods from "ArcFace: Additive Angular Margin Lo..."
...Following the standard evaluation protocol defined for the unrestricted setting [22, 24], we test the model CNNs on 6000 face pairs (3000 matched pairs and 3000 mismatched pairs)....
[...]
...[24] proposed a geometrically interpretable loss function, called ArcFace, which is integrated with different CNN models (e....
[...]
...As seen from Figure 5, both the computed Pearson and Spearman coefficients for the six CNNs on the two datasets are larger than 0.4 in most cases, and particularly, those coefficients for the four more recent CNNs (ArcFace, Light-CNN, SphereFace, and ResNet-Face) are even close to or larger than 0.6 with relatively smaller standard deviations, in agreement with the previous results for VGG-Face1/VGG-Face2 on MultiPIE, which further suggests that the predicted representations by the linear model are strongly correlated with those outputted from CNNs....
[...]
...Addressing the above questions, we investigated the face representation problem at higher CNN layers by formulating face images as points in a parameter space in this work, with six typical multi-layered CNNs for face recognition: VGG-Face [17], DeepID [14], ResNet-Face (defived from ResNet [27]), SphereFace [22], Light-CNN [23], and ArcFace [24], and three commonly used face datasets: Multi-PIE [28], LFW [29], and MegaFace [30]....
[...]
...Deng et al. [24] proposed a geometrically interpretable loss function, called ArcFace, which is integrated with different CNN models (e.g. ResNet [27]) for face recognition and verification....
[...]
2 citations
2 citations
References
123,388 citations
33,597 citations
30,843 citations
13,268 citations
10,447 citations