It seems the code only realizes the location of the target.

Why have you switched to Piotr's toolbox? Has it proved better performance?
I am asking this because the source code of the toolbox is really hard to read and debug.

理论上来讲，gray特征应该比hog快才对，但我在自己的机器上（i9）对dog1这个序列测试时，发现gray明显慢很多：

gaussian + hog 214
polynomial + hog 283
linear + hog 424

gaussian + gray 116
polynomial + gray 122
linear + gray 259

这个大概是什么原因呢

how the update function only return a cv::rect, but how to judge the track failed ?

how to give some adviese?

if (resize_image_) {
pos_ = FloorPointScale(pos_, 0.5);
target_sz_ = FloorSizeScale(target_sz_, 0.5);
}

Here pos_ is halfed but frame is not. This will lead to a wrong crop .

Is there a funtion of file named run_tracker('all') as showed in the picture of running porcess?
So all the videos can be tested.
Hope the author can upload it.