search

tradle / KYCDeepFace

KYC face matching project.
GNU Affero General Public License v3.0
4 stars 0 forks source link

Documentation Questions #11

Open martinheidegger opened 2 years ago

martinheidegger commented 2 years ago

Collection Issue for open questions about this repository

urbien commented 2 years ago

1, 2, 4, 6, 7, 8 - clearly to @fyr91 but the rest are not

I thought the measure of accuracy is already present. Why do you think it is not good?

martinheidegger commented 2 years ago

The first documentation of the lfw_eval and the inference processing I put together here

urbien commented 2 years ago

can you move the documents to github?

martinheidegger commented 2 years ago

Added to the PR #8 in commits https://github.com/tradle/KYCDeepFace/commit/74a1f4128d9bbb4ff346d84d65631ef4040e77a4 and https://github.com/tradle/KYCDeepFace/pull/8/commits/b57289e157605a1ef0f4fb5881a2c702fc5753a3

fyr91 commented 2 years ago

How was the model for the detector trained? DataSource? License?

Face detector are pre-trained model obtained from Link. Its under MIT license

Where does the logic for the face orientation come from? Are there papers or documentation on this approach?

Face orientation is the one we will be facing during the actual usage. When people upload their image, it might not be in the correct orientation. We will use the exif data stored in the image data to determine the orientation of image and correct it accordingly.

Document the structure/source of pairs.txt for lfw_eval.py #12

Lfw dataset is obtained here, within the extracted folder, you will be able to find a pair.txt file for evaluation purpose.

Document the use/function of ArcMarginProduct in the training process.

For ArcMargin please refer to arcface.

Does the code contain copied code from other repositories?

Yes, specifically for recognition training part, it's referred from arcface.

martinheidegger commented 2 years ago

Thank you @fyr91 for filling in the blanks. I have a few follow ups:

Face orientation .... correct it accordingly.

Thank you for explaining how it works, but I was more wondering how did got to this code? Did you think about this yourself? Did you find a paper that described that approach?

Anyways, it seems like this code will be changed to https://github.com/tradle/KYCDeepFace/pull/8#issuecomment-1023517512 which is pretty much self explanatory.

Lfw dataset is obtained here, within

This link returns a 403 - Forbidden?! Is the following a mirror of the same data-set? https://drive.google.com/file/d/1p1wjaqpTh_5RHfJu4vUh8JJCdKwYMHCp/view

Yes, specifically for recognition training part

Just to be clear, this is the only copied code?