zhusz / cvpr15-cfss Goto Github PK

Hello, I downloaded your revised annotation of full AFLW from
http://mmlab.ie.cuhk.edu.hk/projects/compositional/AFLWinfo_release.mat
However, I couldn't find the AFLW-frontal dataset with 1314 images. Could you please help me with the problem? Thank you.

Unconstrained Face Alignment via Cascaded Compositional Learning
你好，请问什么时候能把代码开放呢？
有没有计划开放cfss或Unconstrained的cc++实现呢？

Hi,
In the test mode of singe threading, there is no error when i run the inferenceCFSS.

Because my version is 2014b,
I replace【if (m<100) || (matlabpool('size')==0)】 with 【if (m<100) || isempty(gcp('nocreate'))"】 for using multithreading;
and turn on multithreading by 【parpool(8);】 in the file(inferenceCFSS.m) .
The error as follows:

Thank you!

Hi,thanks for your code.
In your demo video, the algorithm of face detection is also excellent.
What algorithm do you use to detect the face?
Thanks!

I am not sure, I used the model which is offered by this web, but I can't get the result as your paper.
the variable 'er_abs' represents the final error in your inferenceCFSS program, in this vector, from 1 to 330 represents the HELEN test results, the mean value is 13.7559, form 331 to 554 represents the LFPW test results, the mean value is 3.5879, and from 555 to 689 represents the IBUG test results, the mean value is 7.7492. I wonder to know why my results is not the same as yours. Thank you for your explanation.

how to compute 4 landmark in target_simple_face?

The result of the arrayfun() with the imtransform call in transImagesFwd.m (lines 13-14), causes problem when called withimages of much lower dimensions than the new row, col data. Specifically, it fails to specify the spatial locations (x, y) appropriately and thus the program crashes on the extraction of the windows before the vl_hog().

I propose initially a check for the image size, and then a version with imwarp, since as mentioned here it is faster and more efficient.

Therefore, all the lines where you check (matlabpool('size')==0) throw an exception. This can be fixed by replacing them with isempty(gcp('nocreate')).

However, I'm not sure if that call works on older versions of Matlab. If someone confirms this still works then I can make a PR to fix this.

Hi,
When i put vl_sift_toosimple.c into the folder (.\external\vlfeat-0.9.20\toolbox\sift), and then i run the inferenceCFSS.m.
an error occured as followed:
.....
未定义与 'single' 类型的输入参数相对应的函数 'vl_sift_toosimple'。
出错 extractSIFT_toosimple (line 20)
[f,descriptor] = vl_sift_toosimple(img,'frames',fc);
.....

Error using transPoseFwd (line 14)
length(T) must be the same as size(oldPose,1)!

what's wrong?