Clustering of Static-Adaptive Correspondences for Deformable Object Tracking

测试简述：

实际上我测试的时候，有时觉得实时性不好，打开调试开关的时候，显示特征点数量达到3000+，不知道有没有优化的版本，单纯说目标跟踪的性能真的是很好，测试过CT、KCF、LTD，很多人都觉得kcf的性能很惊人，实际上我测试觉得可能kcf最大的特点是实时性，毕竟帧率很高，单纯说目标跟踪的准确性，以我测试的几个视频，可能不太满意。测试的视频motocross.mpg和vtest.avi(opencv自带的)，我觉得cmt的跟踪性能很好,值得大家尝试和改进。

缺点是算法是short-term，遮挡和out-of-view不能很好的解决。

作者首先将OpenTLD的matlab代码转成了c++版本，跟另外一个实现版本，进行过简单的对比，似乎这个版本的实现较优，作者在PHD期间开发的CMT算法。

关于OpenTLD不维护了，感觉比较可惜，解决问题的思路是非常值得学习的，不维护可能是提升空间不大了吧。

作者个人主页：

https://www.gnebehay.com/cmt/

--------以下是作者主页的内容，贴在这里是方便自己查看，如果有不妥之处，请联系本文，谢谢--------

OpenTLD

Disclaimer: This project is now old and will no longer be updated. Please have a look at our new tracker CMT instead.

On this page you can find a C++ implementation of OpenTLD that was originally published in MATLAB by Zdenek Kalal. OpenTLD is used for tracking objects in video streams. What makes this algorithm outstanding is that it does not make use of any training data. This implementation is based solely on open source libraries, meaning that you do not need any commercial products to compile or run it.

The easiest way to get started is to download the precompiled binaries that are available for Windows and Ubuntu. If you have a webcam attached to your PC, you can simply execute tld (on Ubuntu) or tld.exe (on Windows) in order to try it out. For other configuration options, please have a look at the README file. There is also a discussion group of the TLD community where you might get some information. A documentation of the internals as well as other possibly helpful information is contained in this master thesis.

The source code of OpenTLD is published under the terms of the GNU General Public License, so feel free to dig through it. Please understand that this software is meant as a demonstration of what state-of-the-art computer vision algorithms are currently capable of and not as a ready-to-use product. If you find errors in the program, please report them at the GitHub issues page.

CMT

Clustering of Static-Adaptive Correspondences for Deformable Object Tracking

Clustering of Static-Adaptive Correspondences for Deformable Object Tracking (CMT) is an award-winning object tracking algorithm, initially published under the name Consensus-based Tracking and Matching of Keypoints for Object Tracking at the Winter Conference on Applications of Computer Vision 2014, where it received the Best Paper Award. A more detailed paper was published at the Conference on Computer Vision and Pattern Recognition 2015. CMT is able to track a wide variety of object classes in a multitude of scenes without the need of adapting the algorithm to the concrete scenario in any way. Experiments have shown that CMT is able to achieve excellent results on a dataset that is as large as 77 sequences. A C++ implementation (CppMT) is freely available under the BSD license, meaning that you can basically do with the code whatever you want. Additionally, the original Python research code is still available for reference.

How does it work?

The main idea behind CMT is to break down the object of interest into tiny parts, known as keypoints. In each frame, we try to again find the keypoints that were already there in the initial selection of the object of interest. We do this by employing two different kind of methods. First, we track keypoints from the previous frame to the current frame by estimating what is known as its optic flow. Second, we match keypoints globally by comparing their descriptors. As both of these methods are error-prone, we employ a novel way of looking for consensus within the found keypoints by letting each keypoint vote for the object center, as shown in the following image:

The votes are then clustered and outliers are removed:

Based on the remaining keypoints, the new bounding box is computed and the process continues. All the details can be found in our publication.

How to get it?

You can download CMT in either zip or tar formats. You can also browse the source code on GitHub or clone the project directly with Git by running:

$ git clone git://github.com/gnebehay/CppMT

If you use our algorithm in scientific work, please cite our publication

@inproceedings{Nebehay2015CVPR,
    author = {Nebehay, Georg and Pflugfelder, Roman},
    booktitle = {Computer Vision and Pattern Recognition},
    month = jun,
    publisher = {IEEE},
    title = {Clustering of {Static-Adaptive} Correspondences for Deformable Object Tracking},
    year = {2015}
}

Dataset

You can download the dataset (~1.2GB) that we employed for evaluating our algorithm, consisting of 60 sequences collected by Tomas Vojir.

Results

Click here for result videos.