tensorflow_Realtime_Multi-Person_Pose_Estimation
tensorflow_Realtime_Multi-Person_Pose_Estimation copied to clipboard
[QUESTION] image augmentor -- crop range of image -- seems different from openpose
Hi, I'm sorry to disturb you. I have a small question about data augmentation.
Towards def apply_image(self, meta)
of class CropTransform(Transform)
in dataset/augmentors.py, when coordinate of left_up point is negative, it seems that area of original image smaller than 368 * 368 is returned.
For example, suppose that the coordinates of left_up point is (-10,-20), so the valid area of original image will be [0:348, 0:358] according the code, and this area will be given the cropped image, while the rest array value of cropped image will be padded by border_value(128).
But when I read the source code for data transform of openpose-caffe_train in ./src/caffe/cpm_data_transformer.cpp, I find its crop area will be just 368*368 pixels from original image, and not be padded by some value.
Specific code is in function:template<typename Dtype> void CPMDataTransformer<Dtype>::Transform(const Datum& datum, Dtype* transformed_data)
located at line 248.
So could you please tell me whether crop strategy in your code is designed for some special target? Besides, does image padded by border_value helps improve the accuracy?
Thank you very much!