martin-danelljan / continuous-convop Goto Github PK

Hi. In the bottom of the extract_features.m file, the normalization operation is done for the features. I could not understand this normalization operation, can you offer a reference or a website to help me make it clear. And I notice there is no performance degradation when commenting these codes, what role does this operation play? Thanks~

Hi Martin,
In the VOT2016_settings of CCOT, the following line exists.
cnn_params.compressed_dim = [16 64];
Nevertheless, I could not find its usage to compressed the feature maps to the values in the RHS of the above command.
Could you please clarify this point ?

Hi, I wonder is there any demo that can run on feature point tracking as described in the papar?

Hi,

Sorry for posting the irrelevant question in this repo.
I am writing to ask if you are going to release the code and trained models for the recent ICCV19 paper titled "Unsupervised Learning for Real-World Super-Resolution" and if so what is the expected timeline.

Thanks,
Touqeer

when I type the command 'git submodule update'
It has the error:
Cloning into 'external_libs/matconvnet'...
Permission denied (publickey).
fatal: Could not read from remote repository.

Please make sure you have the correct access rights
and the repository exists.

Hi, @martin-danelljan, how to keep the standard inputsize(224*224) to the CNN model in your code? I find that there is a cnndefault mode, but it doesn't work. Thank you very much!

hello, martin,
sorry to disturb you.
Rencenty, I have been doing some work about occlusion and abrupt changes in tracking. I have tested your tracker - C-COT, it works really well. The output is very stable although it doesn't do any thing about tracking - quality - detecting & feedback.
I have also read your paper about tracker - SRDCFdecon, which has done many things about learning rate. The way to set adaptive learning rate has bothered me for a long time, and I have attemped to use PSR and max response . However, it is hard to set a unified threshod which determines tracking state for the whole tracking process. Especially, both of them can't distinguish quick deformation and occlusion. I also attempted to do some improvements based on them, but the result is not so good. I think maybe your tracker - SRDCFdecon can give me some good inspires, and I have visited your homepage about SRDCFdecon. Althogh it says the code will be coming so, I can't wait seeing the wonderful implementation and result. So, I wanna know if it is convenient to share your code with SRDCFdecon.
It will be my plesure to get your reply, thank you !

Hi, martin!
In your CCOT project, dose the reg_filter in tracker.m filer denotes the matrix W_d in the equation (7) as below:

and another question is that why the size of W_d in your paper is (2K_d+2L+1)X(2K_d+1), because in your code, the size of reg_filter is 3x7.

can your take a little time to explain it?

Hi martin
In the model update step of C-COT project, why it adds only half of the new sample "xlf" into "samplesf"?
xlf = cellfun(@(xf) xf(:,1:(size(xf,2)+1)/2,:), xlf, 'uniformoutput', false);

Hello,I copy this template configuration file to my VOT workspace.Then ,it shows that error using system_wrapper>run_once (line 229).It doesn't work,I have tried many times.

Without shifting half an interval T/(2Nd), the interpolated feature map x(t) in the continuous space well agree with the original feature map x[n] in the discrete space:x(0)=x[0],x(T/Nd)=x[1],x(2T/Nd)=x[2],....It is interpreted as a method to preserve the spatial arrangement of the feature pyramid in the paper and the supplementary material. But this operator shift the original feature map 0.5 pixel (T=Nd), and the correspondence between the x(t) and x[n] is not satisfied.

After digged into the tracker.m file in the implementation directory, I found the code a little bit confusing.

According to the ECCV16 paper, the training is done iteratively with CG method and groundtruth. And test is done with gird search and newton method.

But in the tracker.m file, training code is not commented out if the intention is to do test (and the filter is not trained at all). And if the intention is to do training, the groundtruth is not used.

So I wonder if I misunderstood the paper or I just need to tailor the code if I want to use it.

I followed the instructions http://www.votchallenge.net/howto/workspace.html :
(1) Run toolkit_path.
(2) Go to vot-workspace directory. Run matlab and execute workspace_create command.
(3) Select an experiment stack that you want to use. (have prepared sequence)
(4) Enter the unique indentifier of your tracker, e.g. ccot
(5) Select the interpreter used in your tracker, e.g. matlab.
(6) edit the tracker_xxx.m, e.g. tracker_ccot.m

CCOT_repo_path = '/home/wang/projects/vot2017/C-COT/Continuous-ConvOp-master';

tracker_label = 'CCOT';
tracker_command = generate_matlab_command('benchmark_tracker_wrapper(''CCOT'', ''VOT2016_settings'', true)', {[CCOT_repo_path '/VOT_integration/benchmark_wrapper']});
tracker_interpreter = 'matlab';
tracker_trax = false;

(7) I also successfully reproduced your c-cot tracker in another folder, e.g. /home/wang/projects/vot2017/C-COT/Continuous-ConvOp-master

(8)run the script: run_test.m, choose the action 'b', sequence '1', I got some errors:

Error using system_wrapper>run_once (line 229)
No result produced by tracker. Report written to "/home/wang/projects/vot2017/vot-toolkit-master/workspace/logs/ccot/20170511T090124"

Error in system_wrapper (line 39)
[Tr, Tm] = run_once(context.directory, tracker, sequence, start, context);

Error in workspace_test (line 90)
[trajectory, time] = tracker.run(tracker, sequences{current_sequence}, struct('repetition', 1, 'repetitions', 1));

Error in run_test (line 10)
workspace_test(tracker, sequences);

What's wrong with my integration? Thank you!