[Torch vs opencv] image.load vs cv2.load

Recently, I was converting a torch model to mxnet model(The project will be released here) for the convenience of usage on Android OS. The input of the model is an 128*128 image with 1/4 in center is blank(0,0,0) and the output is 64*64. However, when I test the two model with the same image, there are something wrong.
After debugging, I found the input matrix is not the same to the two kind of models. In torch, it calls image.load from package image, while in python, it calls cv2.load from package opencv.

input_torch = image.load('test.png')

input_cv2 = cv2.imread('test.png')

After reading the source code of the functions above, I found three important differences between them.

• The channels order. image.load loads image in RGB; cv2.imread reads image in BGR.
• Data format. image.load organizes pixels in nc*w*h range from 0-1; cv2.imread organizes pixels in w*h*nc range from 0-255.
• Index. For image.load, image matrix starts with 1; For cv2.imread, image matrix starts with 0. It could be more about the difference between lua and python, instead of image.load and cv2.imread.

While, there is also another minor problem might be met. Usually, we want to set the pixels of a area to be a constant value, such as k. Compared with C/C++, python and lua provide convenient methods to do it.
For lua,

input_torch[{{1},{37,92},{37,92}}] = 0

For python,

input_cv2[36:92,36:92,0] = 0

The above means are the same. Since lua and python index from 1 and 0 separately, and lua includes the head(37) and the tail(92) in the change area, while python only includes the head(36) in and the front of the tail(92-1).

That is what I have remembered in this project, please do not hesitate to contact me if you have any problems or something wrong in this po