1、Softmax Loss
official
features
Softmax encourages the separation of features of different categories, but does not limit the distance of feature separation
2、L-Softmax Loss
official
features
- Increase the margin angle parameter m, compress the angle range within the class, achieve compactness within the class, and increase the decision boundary interval between classes
- The parameter m compresses the monotonic interval of the cos function angle, making optimization difficult
3、A-Softmax(SphereFace) Loss
official
features
The weight W is normalized so that ||W||=1, so that the training is more focused on optimizing the depth feature map and feature vector, and reduces the influence of the unbalanced number of samples (the larger the number of IDs, the second norm of weight W The larger the number, W.shape=[embedding_size, cls_nums])
4、AM-Softmax(CosFace) Loss
official
features
- The second norm of the embedding feature x is related to the image quality. The better the quality, the larger the module length of x, so the feature normalization is added, so that ||x||=1, so that the training is more focused on optimizing the angle of the feature vector on, making the obtained deep face features more separated
- After feature normalization, the feature vectors are fixedly mapped to a hypersphere with a radius of 1, which compresses the feature expression space and is not conducive to model convergence. Therefore, the scaling factor s is introduced to expand the feature expression space, which is conducive to model convergence.
- Change multiplicative margin to additive margin
5、ArcFace Loss
official
features
Modified on the basis of AM-Softmax, the margin is changed from the cosine space to the angle space, and the classification limit is directly maximized in the angle space, while CosFace maximizes the classification limit in the cosine space
6、MagFace Loss
official
features
Changed the margin from a constant term m to a function related to the feature modulus length (the larger the modulus length, the cleaner the picture, and more detailed information on the face), and additionally introduced a penalty item based on the modulus length. Quality images impose different penalties (larger modulus lengths have larger angle penalties, smaller modulus lengths have larger modulus length penalties)
7、AdaFace Loss
Face recognition AdaFace study notes_Cassiel_cx's Blog-CSDN Blog
official
features
- The paper explains how the scaling parameter s and margin parameter m affect model training at different values and gives an adaptive parameter adjustment function
- Focus mining and enhancement: hard samples in higher quality images, easy samples in lower quality images, while mitigating attention to difficult samples in lower quality images (if the model is to focus on lower quality images In the difficult sample, the model is likely to learn non-face feature information such as clothing color)
【References】
LOSS of face recognition (on) bzdww
LOSS of face recognition (below) bzdww
Face recognition Loss summary [the most complete mind map] bzdww