mx54039q / fnm Goto Github PK

you have updated model with tensorflow 2

Hi, thank you for your work.

I read your paper and see your test work on various pose. However, I see your code that it only has frontal and profile face.

My question is that Do you prepare various pose code?

If not, How I can do it myself ? do you have any readme for that?

Thanks

Hi,

May I ask why you only used 297,369 images in CASIA-Webface instead of all the images? And I am wondering if you could provide the image list that you used so I can reproduce your results?
Thank you very much!

I downloaded the checkpoint of the pre-trained model, but there is only ck-04 in the file. The error message shows that ck-09 cannot be found. Should I change 09 to 04? Or did you not share?thanks.

按照您之前的建议，我先训练了10轮L1 loss，然后再统一训练，但是还有一点问题，希望得到建议，谢谢。

5轮统一：

10轮统一：

Hi, Thank you for your open code. Can you help me explain preparing the dataset as profile_path, profile_list, front_path, front_list in your code?.

Hi,
Thank you for sharing your great job. I tried to run the training code, but it always leads the computer to restart automatically after a few steps. Have you ever encountered such a situation or do you know what is the problem? Thank you!

Thanks for you offer this work ,I want to get more detail about how to train model? or offer a pre-trained model! thank you!

我在用pytorch复现模型，使用的Multi-pie，但是训练了好多轮了 得到的效果很模糊，loss也很大。。请问您遇到过这种情况吗？谢谢
Loss_G: 24715.967605091395
Loss_D: -28774.34555010077
End of epoch 10

Hi,
Really interesting work, a question relating to the experiments.
The paper shows an impressive boost in performance compared to Light-CNN and VGG-Face, would you still get such boost for VGG-Face2? Or is the model limited to the results of VGG-Face2 as this is the encoder used during training.

Hello, Thanks for your sharing!
Could I ask a question? In discriminator, why you use layer norm instead of batch norm?
Thanks again.

你好，我查看论文的时候发现你使用了MTCNN来对齐人脸，请问你使用的是Facenet里整理的代码吗？如果是的话，你有修改代码里的参数吗？如果修改的话。
minsize = 20 # minimum size of face
threshold = [ 0.6, 0.7, 0.7 ] # three steps's threshold
factor = 0.709 # scale factor
方便告诉我这三个参数的值是多少呢？

最小化yhat和y的差可以理解。然后在加上xhat是出于什么考虑？

Thanks for your great work, I have ran test.py to get the synthesized normalization face from the IJB-A dataset, and verify the recognition perofrmance on the corresponding protocol. However, I found that the performance degrades over 7% in Rank-1 identification when I sent the normalized face into recognition model.

Here's the stepbystep list I have done.
(1) I have use MTCNN's 5-point facial landmark model to get points from input image, and rotate two eye points horizontally. Meanwhile, I set the distance between the midpoint of eyes and the midpoint of mouth with 90 pixels, and crop it into 250x250.
(2) The cropped faces have sent to the face normalization model you provided, and get the normalized faces.
(3) Resized the normalized faces to 144x144 and converted to gray-scale.
(4) Send the normalized faces to Light CNN-29 v2 model to get the facial representation. The Light-CNN model I get is from 'https://github.com/AlfredXiangWu/LightCNN'
(5) Evaluate on IJB-A protocol. (The Light-CNN model only has slight different from paper)

Do you have any suggetion ? I would really appreciate any help.
Besides, do you re-train the Light-CNN model? and which model you have used? could you share the corresponding locations in output image for finding the transformation matrix? Thank you for your attention to this matter, and I look forward to hearing from you.

I am using your repo FNM for state-of-the-art in face normalization. I tried to do the evaluation of FNM (CMU Multi-PIE and IJB-A) on VGGFace and LightCNN again, but the result is not similar your paper. I have some questions on this:

• Which pre-trained LightCNN model you used? 4-layers version, 9-layers version,...?
• On VGGFace, do you need to pre-processing testing images before?
• You are using VGGFace or VGGFace2 for evaluation?

你好，认真研读了，十分精彩。 但是下载pretrain模型时候，达不到预期效果。在实现侧脸变换之后，与regist注册图片相似度比原始侧脸都降低了。 没有找到原因。

我尝试复现你的论文，实验中使用CASIA-WebFace（no-nomal face）和Multi-Pie（normal face ）,训练3个epoch后，利用训练的模型将部分IJBA的人脸数据转正，但效果很不理想。将Multi-pie侧脸转正，感觉效果还不错。我不确定定性实验是否可信，所以请问你是否考虑上传预训练的模型，或者是定量实验的代码？谢谢！！

我发现代码有两处用到了测试test_path，我有两个问题想请教一下：
1、config.py中提到的test_path和test.py中的test_path 是同一个吗？
2、如果他们是同一个，那么我理解的是test_path中图片为不包含训练集的侧脸，如果不是同一个，他们分别指代什么呢？
谢谢！！

As in the prerequisites stands, i created an environment with a 2.7 version of Python (more spesifically 2.7.14) and tried to install Tensorflow 1.8.0. It says:
"""
UnsatisfiableError: The following specifications were found
to be incompatible with the existing python installation in your environment:

Specifications:

• tensorflow==1.8.0 -> python[version='>=3.5,<3.6.0a0|>=3.6,<3.7.0a0']

Your python: python=2.7.14

If python is on the left-most side of the chain, that's the version you've asked for.
When python appears to the right, that indicates that the thing on the left is somehow
not available for the python version you are constrained to. Note that conda will not
change your python version to a different minor version unless you explicitly specify
that.
"""

Should I try a newer Tensorflow version or Python version?

你好，这个想法非常有意思。但是我复现的时候发现，代码中没有用和论文一致的命名。代码里是frontal和profile两个图片输入，是否可以理解为frontal是论文补充材料里面的y，profile是论文补充材料里面的x。

y都是MultPIE的正脸图片（原文：”In constrained experiment, we separate training set of the Multi-PIE Setting-1 into non-normal set (12 poses and 20 illuminations of 150 identities) and normal set (front pose.....“）。 那在unconstrained experiment下，normal set就是frontal pose的MultiPIE了。这样理解对吗？

x都是CASIA中的图片，对应constrained experiment中12个侧脸角度图片。

谢谢！

首先感谢分享这么棒的代码
我试着跑了一下代码，生成的人物正脸照片，相比较而言epoch1的人脸照片会好一些，请问一下，你自己测试的时候，生成的正脸照片效果怎么样？
有一个细节和你的实验不一样，我用的vgg_face 的fc6特征，你可以提供你的npy文件吗

Hello，Author，thank you open source code！About how to train model I want to know the front list and profile list how to bulid,do you have some examples to show?

论文中虽然给了链接: http://cvlab.cse.msu.edu/
project-face-frontalization.html. 但是点进去并没有发现作者给出的网络结构. 所以我想问这个代码的网络结构是你自己想出来的吗?

Hi,

Thank you very much for sharing the awesome work!
For the unnormal dataset, like CASIA-WEBFACE, did you apply alignment to the faces?
Could you please provide the script to preprocess the faces? Thanks!

Hi, I had some problem when I am trying to use your code on MULTIPIE dataset. Could you please provide some demo data for us to reimplement your code?

Thanks a lot~

Hi, that's a nice work and robust model to generate normal face.
I have question about table 2, you said "The performance gap with the other methods lies in two points:

1. These methods fine-tune the baseline (Light-CNN) on the Multi-PIE database, while our FNM is directly incorporated to face recognition model;
2. These methods train with paired data and identity information while our FNM keeps the same training strategy with uncontrolled environment"
   No matter how TP-GAN, PIM and CAPG-GAN are better than FNM on ±90◦, and their fine tune not use VGG-Face, Is that right?
   so, I don't understand why your FNM+VGG-Face is too low almost same as c-CNN and what you would like to represent on Table 2?