Jie Zhang

@mariolew SimilarityTransform is needed for some reason that other papers are already said. And for relative delta shapes, it should be relative not absolute. Imagine that you have a face...

Same implement, not much difference. We need a better way to do data augmentation, No face detector for preprocess.

It seems the program links the wrong dynamic so file. Usually, these so files should lie in `/usr/lib` or `/usr/local/lib`. Error warns you link the files from anaconda.

You can init the shape with mean shape. If you want a data augment, you can also random shift the init shape. Init the shape using other face's ground truth...

Thanks a lot!

@GreenKing 不好意思，代码还没有写完。我也是边看论文边写的代码，同时我也参考了这个项目 [FaceDetect/jointCascade_py](https://github.com/FaceDetect/jointCascade_py) 的代码，这个项目是用 Python 写的，好像训练部分已经写完能够运行，你可以关注一下。 至于移植到单片机上，因为代码用了 OpenCV 的 C++ 接口，可能要用 C 移植起来会比较麻烦但应该不会很难。 速度上的话，论文中提到可以和 VJ 的速度持平，但是论文作者实现的方法肯定是做过优化的，我们自己实现未必能够到达那种速度。

@larsoncs [FaceDetect/jointCascade_py](https://github.com/FaceDetect/jointCascade_py) 并没有使用 OpenCV 做人脸检测，请参考它人脸检测的代码 [FaceDetect/jointCascade_py/demo/demo_detect.py](https://github.com/FaceDetect/jointCascade_py/blob/master/demo/demo_detect.py#L26-L72)

No plan for GPU implementation. As you mentioned this, highly recommend the deep learning method to solve the face detection problem.

Please create a `config.json` file for JDA. `config.template.json` is a template file for configuration. The error is thrown by this [line](https://github.com/luoyetx/JDA/blob/master/src/jda/common.cpp#L117).

try `jda.c` and build with CMake Release mode. `cmake .. -DCMAKE_BUILD_TYPE=Release`