Vot toolkit for python tracker(转)

水深无声 2023-06-24 08:24 44阅读 0赞

issues and possbile solutions to integrate python-based tracker on Ubuntu

I tried to integrate my python-implemented into the vot-2017 evaluation toolkit on Ubuntu 16.04. However, I encountered several issues, and it took much time for me to solve these issues, since I could hardly find any possible solutions from Google.

So, I post these issues and my possible solutions here. Hope this can help you to save your time.

1. Matlab mex compiling, and GCC?

The first time you run the run\_xxx command, the VOT toolkit will try to download (before vot-2017) or compile (vot-2017) the native tools which will be used to accelerate the evaluation or communicate with the tested trackers via the trax protocol. The latest gcc compiler that matlab (R2016b) supported for the mex compiling is GCC 4.9. However, the supported compiler is probably not installed in Ubuntu, in which case you will receive a warning saying that the gcc compiler found is not supported. Fortunately, the mex can still work with the system-provided gcc. But, when you call the mex-compiled function in matlab, you will probably encounter a runtime library error saying that there is something wrong with the standard C runtime library.

This issue is raised because the matlab use different standard C runtime libraries when compiling via mex and running the mex-compiled function. The system-provided GCC is used to compile, but the compiled function is linked to a different GCC runtime library that is bundled with matlab.

My simple solution to this issue is that, we can simply disable the gcc library bundled with matlab by renaming the soft symbol link to the library:

cd MATLAB\_ROOT/sys/os/glnxa64

mv libstdc ++.so.6 libstdc++.so.6.bak

Anytime you find there is something wrong with the library, you can rename it back.

2. Runtime library for your tracker not found?

It is very common for pythoners to run their code in a separate virtual environment. In my case, I installed anaconda (the most popular python distribution for pythoners), and created a separate conda environment for my tracker.

Following the instructions provided by the official support page to integrate your python-based tracker, you will probably encounter import error or runtime-library-not-found error. You need to do a litter more work in addition to what the official tutorial told you.

First, you need to specify your python interpreter in the tracker\_xxx.m file. In my case, it looks like:

tracker\_interperter = ‘path\_to/anaconda3/envs/xxx\_env/bin/python‘;

If your tracker depends on some other library that is outside the conda environment, you also need to specify it. In my case, my tracker is dependent on the CUDA library. I need to specify it in tracker\_xxx.m file:

tracker\_linkpath = \{‘/usr/local/cuda/lib64’\};

3. trax tool do not support python3

The original trax tool for python in vot-toolkit only supports python2. You need to make some changes to the trax module in the native directory to support python3-based tracker. Fortunately, you only need two steps.

First, you need to replace all the ‘xrange’ found in the native trax module files (located at native/trax/python/\*.py) with ‘range’.

Second, replace line 70, file ‘trax/region.py’ with ‘tokens = list(map(float, string.split(‘.’)))’.

[http://www.votchallenge.net/howto/integration\_channels.html][http_www.votchallenge.net_howto_integration_channels.html]

## Python ##

上一步的 *MATLAB* 可以正常运行，说明 *TraX* 库正常，就可以测试官方提供的 *python tracker*. 可以直接对上一步*workspace* 中的文件进行修改，没必要重新创建 *workspace* (特意选择Python)。

还是修改 *workspace* 下的 *tracker\_NCC.m* 文件，修改为：

<table> 
 <tbody> 
  <tr> 
   <td> <pre>1
2
3
</pre> </td> 
   <td> <pre>tracker_label = ['NCC'];
tracker_command = generate_python_command('python_ncc',{'/home/ubuntu/VOT/vot-toolkit/tracker/examples/python'});
tracker_interperter = 'python';
</pre> </td> 
  </tr> 
 </tbody> 
</table>

将 `generate_matlab_command()` 修改为 `generate_python_command()`，参数还是对应 *python* 版的 *tracker* 文件名和目录，`tracker_interperter`其实不用修改，但是给他点面子改成 `'python'`.

运行 *run\_test.m* 试一试，果然出错了！当我 *Anaconda* 不存在吗？

*  **wrong python interpreter**
    
    <table> 
     <tbody> 
      <tr> 
       <td> <pre>1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    </pre> </td> 
       <td> <pre>...
    CLIENT: Creating process /usr/bin/python  -c "import sys; sys.path.append('/home/ubuntu/VOT/vot-toolkit/native/trax/python');sys.path.append('/home/ubuntu/VOT/vot-toolkit/tracker/examples/python'); import python_ncc"
    CLIENT: Working directory is /tmp/tp8b02cd23_8b86_416c_8e82_000ea94a4639
    ...
    Traceback (most recent call last):
      File "&lt;string&gt;", line 1, in &lt;module&gt;
      File "/home/ubuntu/VOT/vot-toolkit/tracker/examples/python/python_ncc.py", line 6, in &lt;module&gt;
        import cv2
    ImportError: No module named cv2
    ...
    </pre> </td> 
      </tr> 
     </tbody> 
    </table>
    
    但是官方并没有说如何选择 *python* 解释器。谷歌一下，有[线索][Link 1]，根据这条线索，修改`tracker_interperter`：
    
    <table> 
     <tbody> 
      <tr> 
       <td> <pre>1
    </pre> </td> 
       <td> <pre>tracker_interpreter = '/home/ubuntu/anaconda2/envs/tf15/bin/python';
    </pre> </td> 
      </tr> 
     </tbody> 
    </table>
    
    还是报同样的错，解释器仍然是`/usr/bin/python`（早就说 *tracker\_interpreter* 没用了！）。
    
    也许是版本原因造成这种方法不再适用，但是，观察 *generate\_python\_command.m* 文件：
    
    <table> 
     <tbody> 
      <tr> 
       <td> <pre>1
    2
    3
    4
    5
    6
    7
    8
    9
    10
    11
    </pre> </td> 
       <td> <pre>function command = generate_python_command(script, paths)
    
    trax_python = get_global_variable('trax_python');
    python_exec = get_global_variable('python');
    
    % If path to python trax implementatin is set then we attempt to export it to tracker
    if ~isempty(trax_python)
        paths = cat(1, {trax_python}, paths);
    end
    
    ...
    </pre> </td> 
      </tr> 
     </tbody> 
    </table>
    
    `python_exec`不就是解释器么，编程规范真是好东西。修改它肯定可以解决问题。在 *workspace* 下的 *configuration.m* 文件中添加一句：
    
    <table> 
     <tbody> 
      <tr> 
       <td> <pre>1
    </pre> </td> 
       <td> <pre>set_global_variable('python', '/home/ubuntu/anaconda2/envs/tf15/bin/python');
    </pre> </td> 
      </tr> 
     </tbody> 
    </table>
    
    成功运行。

## **Conclusion** ##

VOT toolkit 调用 *python* 版本的 *tracker* 需要修改 *workspace* 目录下的两个文件

1.  *tracker\_NCC.m* 中：
    
     *  `tracker_command` 的函数名和参数
     *  `tracker_interperter` 改为`'python'`（好像没什么用）
     *  如果 *tracker* 使用了 *CUDA、cuDNN*，需要加上`tracker_linkpath = {'/usr/local/cuda/lib64','/usr/local/cuda/extras/CUPTI/lib64'}`
2.  *configuration.m* 中：
    
     *  添加 *python* 解释器路径：`set_global_variable('python', '<interpreter_path>');`

由于 *Trax* 支持了 *Python3*（*VOT/trax-master/python/trax* 目录下的 *python* 文件中有以下语句）

<table> 
 <tbody> 
  <tr> 
   <td> <pre>1
2
3
4
</pre> </td> 
   <td> <pre>...
if (sys.version_info &gt; (3, 0)):
    xrange = range
...
</pre> </td> 
  </tr> 
 </tbody> 
</table>

所以不必特意将 *python demo* 中的 `xrange` 改成 `range` 。

[http_www.votchallenge.net_howto_integration_channels.html]: http://www.votchallenge.net/howto/integration_channels.html
[Link 1]: https://groups.google.com/forum/#!topic/votchallenge-help/9RmJ61uXImo