tencentyouturesearch / adv-makeup Goto Github PK

Hi.
We think that the th_dict in the test.py may be the dict of thresholds for different FR models. However, we don't know the calculation method of these thresholds in the dict exactly. Could you tell us is th_dict [name_of_model][i] the threshold of name_of_model at 1e-(i+1) FAR in the LFW dataset?
Thanks.

Hi, thank you for sharing the implementation of your great work! And I have some concerns with the following line of code:

Why do you clip the grad of the discriminator when updating the generator (maybe instead of clipping the grad of generator)? I fail to find the corresponding explanation in your paper.

I also notice that when updating the discriminator, fake images are not detached from the generator network (which is a common practice to stop the grad from being propagated to the generator). Do you intend to do so?

I really appreciate it if you could kindly solve my issues.

Best Regards

Hi, in paper it has been mentioned that the models will be selected randomly during meta learning (training phase). However, in model.py file, in line 264, only the first 2 models is selected every time as meta train models. Should not it be random from 3 models?

您好，我想在lfw数据集上复现adv-makeup，请问可否提供lfw数据集的landmark的pickle文件？

研发人员您好，在根据论文中的实验设置描述复现文中的数字攻击的实验一时对实验一中的实验设置有一些疑惑，想请教以下问题:
1.LFW与makeup数据集的单次实验的攻击目标图像都是从LFW中选出的1000张图片中随机选择10张图片进行攻击吗？
2.论文表格1所记录的数值结果都是针对LFW数据集中所选出的1000个目标的平均测试攻击成功率吗？

Thank you for open-sourcing this.

Because there is no landmark detection code here, it's difficult to directly apply this attack with real images.

The eye regions are created by indexing the landmarks array. But there are no instructions on the correspondence between the index and the physical meaning of the point. This makes utilizing other existing landmark detection methods difficult.

Could you tell me the relation between the index and the meaning of the point? For example, what does the 94th point in the landmarks mean, left_eye_left_corner, right_eye_top, or something else?

Thank you very much.

Hi! After we run python train.py, an error occurred:

Traceback (most recent call last):
  File "train.py", line 3, in <module>
    from model import *
  File "/foo/Adv-Makeup/model.py", line 14, in <module>
    from encoder import *
ModuleNotFoundError: No module named 'encoder'

 
Could the original authors fix the bug?
Thanks a lot.

为什么经过mtcnn对图像对齐和裁剪后，600*600还是报错，
File "/home/yuanpeng/桌面/Adv-Makeup-main/dataset.py", line 47, in getitem
before_lmks = self.api_landmarks[before_path].astype(np.int)
KeyError: 'before_aligned_600/00288_0_0.jpg'

请问这篇论文中的conf置信度实验数据，代码中哪里实现了

Thanks.

Hello.

Thank you for your work! Could you please provide some pretrained models for making experiments with adversarial makeup. Unfortunately I have no opportunity to train it on myself.

研发人员您好，非常感谢您开源Adv-makeup这一项目，关于论文中物理实验的部分，我有一些问题想请教：
1.请问您所承载对抗眼影的纹身贴纸是自己打印的还是第三方纹身贴纸店铺代为制作？如果是后者能否提供一下店铺信息？
2.请问纹身贴纸是直接按照数字图像修改的区域直接使用，还是在得到原板后对要实际使用区域进行裁剪？
3.在查看您的实验结果时发现您的物理妆容离眼镜边缘较远，而数字结果的妆容一直画到了眼白与眼眶的边缘，请问是您是在物理实现时修改了mask改变了对应数字对抗样本的化妆区域还是其他？
我承诺我的物理条件复现实验只在实验室环境内进行，不会使用在任何实际场景，希望您能不吝赐教。
再次感谢您进行这项有意义的工作并将其开源。

您好，我在代码中似乎没有看到论文中内容损失的实现？并且多了一个用于去噪的tv_loss，二者的作用是相似的吗？

您好，下面两张图的最后一列图像是我使用开源项目针对对应目标生成的妆容对抗样本。我发现相对于论文中展示的结果，我所生成的妆容对抗样本Face++的置信度普遍偏低。有两张图像的置信度差距较为明显，请问是将对抗样本上传到face++时进行了预处理？还是一些细节出现问题？是否能提供face++检测对应的代码？

各位研发人员您好，以下我通过本项目代码以及readme中提供的特征点文件、数据集与模型参数文件的部分生成图像。生成图像在粘贴的部分与原图存在较为明显的颜色差别，自然程度相较于论文结果差一些。请问产生这种现象是否为正常现象或者是可能是代码的使用上哪里出现了问题？

非常感谢大佬的开源：）

想请问一下，您在paste_patch函数里定义了before_orig, before_pasted, target_img, fake_afters, before_pasted_eyes这五个数组，请问分别有什么含义呀？

由于在复现论文实验效果中遇到一些问题，想请教各位研发人员以下一些问题:
1.请问在使用LFW数据集进行训练时，用于输入判别器的妆容图像是使用的makeup数据集中妆容图像吗？
2.使用makeup数据集进行训练的素颜图像是数据集中剔除100张与10张作为有目标攻击的图片后剩下的图片吗？
2.论文中选择出的198张妆容图像是如何选择的？

Thanks.

作者您好，我对您的这个项目很感兴趣，在阅读过程中发现您的项目结构中为了让生成的眼影更加自然的覆盖在人脸上使用了后期融合的处理，请问处理完后的对抗样本还能具有原本的攻击效果么？

Hello!

Thank you for great work!
I would like to know if it is possible to conduct experiments with trained models on your own images. If so, how to generate landmarks?

您好，我想要在当前模型的基础上增加预训练的人脸识别模型观察攻击效果是否提升，想请问一下论文中各个人脸识别模型的FAR的计算方式，时使用的什么数据集进行计算的？数据量是多少？

你好，
麻烦问一下，如何获取Pretrained_FR_Models 这个模块，这个模块无法找到。查询网上的资料也没有获取到相关信息，是私有模块吗？

各位项目研发人员您好，非常感谢您将该项目代码公开，由于个人计算资源有限，我无法在自己的设备上进行训练。因此，如果方便的话想请您提供一下预训练的模型参数，谢谢。

您好，我想在lfw数据集上复现adv-makeup，以及根据LADN生成更多对抗样本，请问数据集的landmark的pickle文件是怎样生成的？我观察到您把LADN的图片都裁剪到600x600，请问是怎样裁剪的？谢谢

请问论文中物理对抗的贴纸是使用的什么材料呢？看起来不像是纸张。期待回答，谢谢！

We would like to generate the landmarks of our face images.

We see that a numpy.ndarray in landmark_aligned_600.pk contains 106 2D points. However, we can only generate 5 2D points using mtcnn. Could the authors add the code of generating the landmarks in the repo or describe how to generate the landmarks?

Thanks.