jponttuset / mcg Goto Github PK

i download the code in my ubuntu environment.
i want to reproduce the segmentation result in paper, so i choose the pre-trained document
first, i run install.m eveything is ok.
then, i run the demo_im2mcg.m, something happend as following:

Error using eigs
Too many output arguments.

Error in spectralPb_fast>ncuts2 (line 130)
[EV, EVal, debug1, debug2] = eigs((D - A) + (10^-10) * speye(size(D)), D, n_ev, 'sm', opts);

Error in spectralPb_fast>ncuts_downsample3 (line 107)
[EV, EVal] = ncuts2(A_down, NVEC);

Error in spectralPb_fast (line 40)
[EigVect, EVal] = ncuts_downsample3(W, nvec, 2, 2, [ty, tx]);

Error in img2ucms>img2ucm (line 110)
[ sPb_thin] = spectralPb_fast(owt2 * mult_Pb, nvec, ic_gamma, dthresh) / sat_sPb;

Error in img2ucms>img2ucm_scale (line 86)
[ucm2, times] = img2ucm(I, model, param.mult_Pb, param.sat_sPb, param.nvec, param.dthresh,
param.ic_gamma);

Error in img2ucms (line 67)
[ucms{s}, times] = img2ucm_scale(I, model, param);

Error in im2mcg (line 73)
[ucm2,~,times] = img2ucms(image, sf_model, scales);

Error in demo_im2mcg (line 10)
[candidates_scg, ucm2_scg] = im2mcg(I,'fast');

it look like something wrong inside the eig function

thank you!

Hi, I would like to thank you for sharing the code with us.
I have a question, I have downloaded MCG Proposals for PASCAL VOC 2012 in here and in each .mat file there are

• superpixels
• labels
• scores
  I visualize the superpixels and it looks like this
  
  I wonder what should I do to obtain segmentation like the rightmost column in Fig. 2 of MCG paper.
  

Thank you!

Is any windows version of complied mex files?

dear author,
I download the BSDSdatasets MCG Proposals in link:
https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/mcg/#results

the file in each mat including:

1. superpixels:481*321 uint16
2. scores: 6340*1 double
3. labels: 6340*1 cell

i have try only using the superpixels for evaluation in BSDS500 benchmark here:
https://github.com/BIDS/BSDS500

but i get the result like this
GT=0.35, RI=0.77, VI=4.24

in the paper, the report result in page5 table 1 is:
SC=0.611, RI=0.813, VI=1.548.

i want to know if correct for my operation, maybe i have some mistakes..

thank for your reply!

Hi,
I encounter the following error when I run build.m in octave:

>> build
mkoctfile: unrecognized argument /home/SENSETIME/lintsuihin/mcg/pre-trained/lib
warning: mkoctfile exited with failure status
warning: called from
    mkoctfile at line 171 column 5
    mex at line 29 column 18
    build at line 66 column 5
mkoctfile: unrecognized argument /home/SENSETIME/lintsuihin/mcg/pre-trained/lib
warning: mkoctfile exited with failure status

Any help would be much appreciated, thanks!!

in windows system, the "aux" is a special directory name，so we should not use this to be directory name when want to compile the cpp code into matlab

For very flat images, if there's a UCM with no contours, the code breaks.

The candidates score at the output is the one obtained by the random forest regressor, not the one after the max margin is applied. This way, the proposals are not sorted by the available score, and this can be misleading.

We should save both scores to avoid any confusion.

Very big images (bigger than full HD) make the code crash. Not sure if it's the contour detector though, I should do a performance evaluation with respect to the size of the image, and individualized at each main part of the code.

Does this method require the full semantic segmentation per-pixel labels or the instance segmentation per-pixel labels ?

Thanks!

Dear author:
I can't find the 'results' folder in https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/mcg/#results.
Precomputed results not found: Have you downloaded them? Visit the MCG website and download the 'results' folder.

How can I use this code to group my own generated proposals? Which script should I use?

I currently survey various segment proposal methods.

However, I could not catch what kinds of ground-truth labels (e.g., segmentation map, object boundary) are required for MCG training because I'm not familiar with Matlab code :(

Also, some papers argue that MCG is an unsupervised method, is it right?

If anyone knows about this, please let me know.

root_dir has to be set manually, we can make it work automatically

Can you share the pretrained model for PASCAL Dataset?

Hello,
Is there a C++ version of this code? if not, do you know some method to compute UCMs with python or C++ instead of matlab?Thanks!

When compiling with certain versions of GCC and Matlab, a lot of warnings come up regarding redefinitions of constants in cmath, math.h, or math.hh.

It compiles and runs fine, but it's annoying.

Is there a Python version of this code ? a matlab license is very costly! Thanks!

The files are that ended with mexa64 compiled with linux，do you have the windows version. Apperciate it.

For the proposals generated for COCO, was MCG pretrained only on PASCAL and BSD? or was it pretrained on COCO?

Thanks!
--Issam

what matlab version you use ?

The algorithm has some kind of randomness that, although does not affect the quality significantly, makes the results not reproducible.

Dear author：
Thank you very much for the open source code, I still have some questions to ask.
Can I re-train MCG on a new dataset (e.g., natural scene text images) to get proposals or segmentation results?
Looking forward to your reply!

Hello, thank you for your code and it's really helpful for me. But when I run the code of demo_im2ucm/demo_im2mcg in full folder, it gets one error is 'Reference to non-existent field 'opts'', I found that it happens in edgesDetect.m, I can not understand why, so I'm looking forward to your answer. thanks~

Hi,
I'm not sure if it's appropriate place to ask, but I wonder what dataset was used to train the model sf_modelFinal.mat?

Thanks in advance,
Nikita

Hello! Your work is very helpful for my research. When i implement this demo file, I find a question which bewilder me.

I get the candidate result and bbox from the following way, but sometimes the candidate result and bbox are mismatched.

mask = ismenber(candidates.mcg.superpixels, candidates_mcg.labels{id});
box = candidates_mcg.bboxes(id);

There is another way to get the matched candidate result and bbox?

Looking forward to your answer!

Dear author:
there's a trouble about the function 'loadvar()' in the 'im2mcg.m' line 52. The compiler always told me that 'can't find the 'loadvar' function'. so, i believe there should be several reasons. First, i didn't load some libraries or add some folder paths. Second, the function is lost in my download.
I hope you can give me some tips when your time is convenient.
Yours Sincerely.

when I run the file of install.m. I got the error of The needed function (mex_assess_one_sel) not found. Have you built the package properly? but I have already run the file of build.m.