cleardusk / meglass

An eyeglass face dataset collected and cleaned for face recognition evaluation, CCBR 2018.

Hi there, I'm Jianzhu Guo 👋

meglass's Issues

Error with data distribution

The reported number of glasses and non-glasses images is reversed. There should be 14832 glasses and 33085 non-glasses. Can you guys verify this? Thanks.

The specific method of eyeglass faces synthesis in your paper?

I am trying your method to improve the performance of my face recognition model.But I can't understand clearly，How are these eyeglass faces generated concretely？How do you blend 'Synthetic Result' from 'Input Face Image' and '3DMM face model' ? Can you give me some advice? Thx!

Glassed detection and location

Excuse me ,im a PhD student and now solving the glassed detection problem.
Thank you very much for you opened MeGlass dataset.
And will you open you code in your paper later?

For the same identity, how can i judge the images with or without eyeglasses?

Excuse me, this file, named MeGlass_120x120.zip, only contains 47917 images. And according to the naming rule, I can just distinguish the face image's identity. But how can I judge the images with or without eyeglasses for the same identity?

用3DMM去合成不会造成人脸变形吗？

Can you share your code in your paper?

issue for eyeglass generator

i want generate eyeglasses for my own face dataset, how could i achieve it? by this project? any related or useful eyeglasses generator used?

Evaluation protocol IV get the different result from paper

Hi
I evaluation the four testing protocols in paper. I don't confirm my method and result is correct.
there is 4 file list txt。[0] gallery_black_glass.txt [1]gallery_no_glass.txt [2]probe_black_glass.txt [3]probe_no_glass.txt
4 test protocols:
I) [1] vs [3]
II) [0] vs [2]
III) [1] vs [2]
IV) [0][1] vs [2][3] (Gallery images contain both eyeglass images and non eyeglass images, so as probe image.)
the result from paper show that III) has the better result than [IV)。
for ResNet-22-A RPR@FAR=10e-4 III): 88.13 IV):78.17 Rank1 III): 95.61 IV):92.31

in my experiment test, I use LightCNN29 model to evaluate, ang get the following result:
for LightCNN29 RPR@FAR=10e-4 III): 87.49 IV):87.03 Rank1 III): 93.39 IV):98.54
as showed, III) has litter better result than IV) and Rank1 smaller than IV）.

I mixture [0][1] as the new Gallery, and mixture [2][3] as the new probe in protocol IV). which should has less degree of difficulty than protocol III), because all of pairs in III) contains non-eyeglass and eyeglass, otherwise，IV)contains some simple pairs of non-eyeglass.

So I don't think IV) could get better than III) from my consider and test

where is wrong for my test? could you tell me

Downloading

Hi there,

Would it be possible to create a download link that is not Baidu Yun as I am struggling to download the data and test the dataset.

Thanks,
Asim

segmentation masks?

Do you have the ground truth for segmentation masks of where the glasses are on each image?
Specifically, I would like to know the ground-truth masks for pixels that correspond to the glass of the eyeglasses (I'm building an eyeglass segmentation network).

generate eyeglasses 3d module

how do you generate 3D data of the glasses ?

Reminder: release the code please

@cleardusk Hi, can you release the code on your technique which would help a lot of mates make their own datasets ? Cheers!

您好，近期还有开源代码的计划吗

eyeglasses reconstruction

your work is very excellent, can you release your code as describe in paper?

Could you share the original MeGlass data on Google Drive?

Hi @cleardusk, thanks for sharing your data & codes. Could you share the original MeGlass data on Google Drive? It's hardly possible to download it from BaiduYun.

Labels

Instead of a separate text file, rename the images with the label itself