Generating Feature Embeddings using ArcFace+ResNet100

[sohaibjawad]

Hi! I am following your repo as is, but the only difference is that I am computing feature embeddings using ArcFace+Resnet100 pre-trained model. Could you please guide me through the steps you took to generate .mat files of feature embeddings? I am using the face bounding box information given in train split files and then computing it's embedding. The TAR@FAR values I am getting are significantly lower than yours. Waiting for your response!

Thanks

[YirongMao]

Hi! I am following your repo as is, but the only difference is that I am computing feature embeddings using ArcFace+Resnet100 pre-trained model. Could you please guide me through the steps you took to generate .mat files of feature embeddings? I am using the face bounding box information given in train split files and then computing it's embedding. The TAR@FAR values I am getting are significantly lower than yours. Waiting for your response!

Thanks

My face model is ResNet-34, trained on WebFace with softmax loss. If your face model is robust enough, NAN may be unable to improve the performance.

[sohaibjawad]

I see. And did you preprocess the images like image alignment etc? If yes which technique did you use?

…

On Mon, 8 Feb 2021, 07:49 Yirong, ***@***.***> wrote: Hi! I am following your repo as is, but the only difference is that I am computing feature embeddings using ArcFace+Resnet100 pre-trained model. Could you please guide me through the steps you took to generate .mat files of feature embeddings? I am using the face bounding box information given in train split files and then computing it's embedding. The ***@***.*** values I am getting are significantly lower than yours. Waiting for your response! Thanks My face model is ResNet-34, trained on WebFace with softmax loss. If your face model is robust enough, NAN may be unable to improve the performance. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQN2YBBF537ZDATNWMK32ATS55GNHANCNFSM4XHV5J2A> .

[YirongMao]

I see. And did you preprocess the images like image alignment etc? If yes which technique did you use?
…

I only detect face with MTCNN(https://github.com/ipazc/mtcnn). No other preprocess steps.

[sohaibjawad]

Okay thanks I'll give it a try

…

On Mon, 8 Feb 2021, 08:12 Yirong, ***@***.***> wrote: I see. And did you preprocess the images like image alignment etc? If yes which technique did you use? … <#m_-2609376644158029007_> I only detect face with MTCNN(https://github.com/ipazc/mtcnn). No other preprocess steps. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AQN2YBHGVC4B7QXCBDESX63S55JBZANCNFSM4XHV5J2A> .

[sohaibjawad]

How did you compute feature embeddings using mtcnn for subjects where there are more than 1 person in the frame?

[tuantran23012000]

Hi! I am following your repo as is, but the only difference is that I am computing feature embeddings using ArcFace+Resnet100 pre-trained model. Could you please guide me through the steps you took to generate .mat files of feature embeddings? I am using the face bounding box information given in train split files and then computing it's embedding. The TAR@FAR values I am getting are significantly lower than yours. Waiting for your response!

Thanks

Can you provide IJBA dataset?

[hairenaa]

我运行了您的代码但accuracy显示很小，甚至是0，这是为什么呢？，另外我并不能清晰地看到原始数据，请问您有IJB-A的数据集吗？
I ran your code but the accuracy shows very little, even 0, why is that? , In addition, I can't see the raw data clearly, do you have an IJB-A dataset