峰哥，位姿参数优化求导的部分有BUG，所以你的跟踪效果很差，我也是19年12月才把这个BUG改好，这个BUG拖了我2年啊😓。具体就不说了，我毕业后也准备放代码。

Hi,

Thank you for sharing your great work. If you do me a favor, I would like to ask you the test data, cube.avi and obj files.

Thanks,
Bobby

程序运行前，要输入首帧的pose，请问这个pose如何得到？

I have finally found the time to integrate this into OpenCV. See:
https://github.com/opencv/opencv_contrib/tree/master/modules/rapid

Here, I ended up rewriting most of the code, but having your code for comparison was a nice help nevertheless.

I found some bugs, which I wanted to note though (Off the top of my head)

• sometimes the search-lines were reversed, which led to mismatches
• your LOS implementation was using the background class for tracking
• you were initialising memory very inefficiently, which hampered runtime

Fixing these allows using the non-robust OpenCV pose-estimator instead of levmar with tuckey.
The version that is now in OpenCV is about 2-3x faster, while being more precise.

On the plus side, the non-robust estimator showcases the difference between GOS and LOS better.

what are the licensing terms of this? I would like to port it in OpenCV contrib, however it must be BSD2 licensed for this.