different face pairs of LFW has high similarity, why?

[xiakj]

environment: (1) retinaface(R50) for face detection and alignment (2) arcface(r50) for face feature extraction (3) cosine similarity

I test LFW face pairs using arcface, the same face pairs has higher cosine similarity, around 0.8~0.9, but different face pairs also has high cosine similarity, around 0.6, i think it is abnormal, but i can not make out the issue. can you share me with some idea?

[liutianyuan]

environment: (1) retinaface(R50) for face detection and alignment (2) arcface(r50) for face feature extraction (3) cosine similarity

I test LFW face pairs using arcface, the same face pairs has higher cosine similarity, around 0.8~0.9, but different face pairs also has high cosine similarity, around 0.6, i think it is abnormal, but i can not make out the issue. can you share me with some idea?

Have you solved this issue? @xiakj

[xiakj]

@liutianyuan 算是解决了吧，是因为我计算cosine相似度的程序写错了。

[liutianyuan]

@liutianyuan 算是解决了吧，是因为我计算cosine相似度的程序写错了。

@xiakj 哦哦，了解。我用的sphereface的模型。我这边是抓拍的侧脸图片容易误识别，我看了下，侧脸和有的底库照片cosine相似度很高，有的很低，也没啥规律。您这边有啥建议没。谢谢。

[xiakj]

@liutianyuan 我的也是，模糊图像和底库相似度也挺高。也不知道该怎么解决，目前是把低置信度的人脸框、侧脸过滤掉不处理了，来降低一点风险。

[mzxhzhp]

@liutianyuan 我的也是，模糊图像和底库相似度也挺高。也不知道该怎么解决，目前是把低置信度的人脸框、侧脸过滤掉不处理了，来降低一点风险。

看到这里我有个疑问，底库是怎么生成的呢？道友可否指点一二，感谢。

[xiakj]

@mzxhzhp 什么意思，底库不就是图像的特征值么

[mzxhzhp]

@mzxhzhp 什么意思，底库不就是图像的特征值么

好吧，懂了。是我想复杂了