taylover-pei / ssdg-cvpr2020 Goto Github PK

I see that all the datasets are available in Video format. Is there any code/pipeline you use to process the database into images ?

Hi, i'm trying to train SSDG to another dataset.
I trained Oulu+SiW+ReplayAttack(which i could acquire) and tested another dataset.
The best results are achieved in just 2 epochs.. and i think it is a little bit odd..
So i want to check your training log. Could you share?

你好，是否可以分享一下热力图的代码呢

我现在是每一张图片一个，但好像是有问题，您能不能分享一下初始4个集的list形式？
就是generate_label.py需要的格式

Can you share your trained model, if possible?
I can see only a standard resnet18 model in the folder 'pretrained_model'

Hi, i have researched face anti spoofing area.
I got Oulu-NPU & ReplayAttack dataset, but i couldn't get MSU-MFSD & CASIA-MFSD dataset.
In the case of MSU-MFSD dataset, i send agreement several times but there is no reply...
In the case of CASIA-MFSD dataset, i want to upload agreement on website(http://www.cbsr.ia.ac.cn/images/xyz/AgreementUpload.aspx) but failed every time with no reasons.
So, Can you tell me how to get the dataset?(MSU-MFSD, CASIA-MFSD)

Hi! First of all, thanks for the code! It is very helpful. I tried to reproduce the tSNE features from the embeddings but had no luck. The patterns are logical, though, but don't know which parameters did u use or from which layer did u extract the embeddings to produce the figure 5 from the paper. Any help would be much appreciated. Thanks in advance!

Hi, is it possible to share pre-trained weights for complete model.
Thank you.

single_video_frame_list[6 + j * frame_interval])
Does the above code mean we choose the fixed(固定的) one or two frames from each video,
and the rest frames of each video are abandoned?

Can I replace 6 with other number，such as 1,2,3?
Looking forward to your reply.

第一个问题：请问预处理的流程是不是将每个视频的所有帧都提取出来保存然后再训练，
还是说一边提取一边训练；

第二个问题：比如一个视频10秒钟，30帧每秒的话，总共就是300帧
我需要都保存还是只保存一部分。

非常希望能够得到您的回复

主要想看一下您的train_list的格式

请问对数据集的预处理就是将视频的每一帧抽出，利用MTCNN处理成256X256X3的png格式么

Hi. I think the weight normalization should be along the dim of 1 since the size of weight is (2, 512). But in the code, the weight normalization is along the dim of 0.

when calculating Real_AdLoss of real face , how to set the adversarial label , Why does it have 2

Figure 4. Grad-CAM [32] visualizations in your paper show some of the input images.
I notice that you used not only the face map but part of the background, could you please tell me the detail about data processing.When I crop the face with detected bbox, no background will be left.
Thx for your great job~

How did you draw the ROC curve? If it's convenient, could you share the comparison method and your FAR & FRR data? Thank you very much!

Hi,
I would like to re-train model but the size of all dataset is too much with me. Could you share pre-trained model for us.
Thanks in advance.

In your batch hardest triplet loss code, the margin is hard-coded with 0.1.
https://github.com/taylover-pei/SSDG-CVPR2020/blob/master/loss/hard_triplet_loss.py#L50

the value 0.1 should be replaced with self.margin.

作者您好，我想问一下你的可视化部分中的t-SNE是怎么做出来的？能分享一下代码吗？

Inspiring paper!
Many thanks for sharing the code.

I wonder which datasets you trained the released model.
Shouldn't four experiments have four pretrained models.

Hi there, I achieved different results with your current config files and three of them are lower performed.

Mine was like this

I just wonder if Hyperparameters in config are the ones achieved results in the paper? Many thanks

I don't see folder config in your repo. Please help me

How should the pretrained model be loaded? I loaded the resnet18-5c106cde.pth model provided in the pretrained_model directory into the torchvision resnet18 model. This model has an fc layer with 1000 outputs. I used the first output as the 'spoof' class score and the second output as the 'real' class score. I evaluated the model on OULU but the accuracy is very bad. Am I loading the pretrained model correctly?

Because of a "runtimeError: Legacy autograd function with non-static forward method is deprecated." I had to change the class GRL but now I have another error regarding the variable "iter_num" since I cannot initialize it with the static style...Any help?
DGFAS.zip

File "../../models/DGFAS.py", line 219, in forward
ctx.iter_num = ctx.iter_num + 1
AttributeError: 'GRLBackward' object has no attribute 'iter_num'

Any idea how to obtain the CASIA dataset ? The publicly available link seems to be broken.

Is one gtx1080ti with 11GB enough to run this code successsfully?

hello!
Since the MTCNN in the given link is for face detection on images, is the input of MTCNN in this experiment all the frames of all the videos in the dataset? Because I see that sample_frames is extracting one or two frames from many images（ are the images from all frames of a video?）, in that case why not extract one or two frames in advance and put them into MTCNN for detection?
Looking forward to your reply.

Hi. As mentioned in paper you normalise feature using l2 norm. But in your code you devide features by square root of norm multiplied by √2. Can you share your idea behind this approach ?

why the triplet loss can not descend until margin value 0.1

您好，感谢分享的作品，请问在人脸对齐部分，猜测您可能使用了相似变换或者仿射变化，那预先设定的五个人脸关键点位置是多少，能分享下吗

Hi,

First of all, thank you for providing an open-source implementation for your interesting paper, what is the pipeline to get an inference of your model? I want to test your model but I don't know what is the instruction to do that.

Thanks in advance.

hello @taylover-pei,
I'm a student, i have a project for anti spoofing. I have found out through your method, it's great!!!
But my computer is not enough to train the model.
Would you like to share with me!
Thank you so much
Contact with me: [email protected]