Different mean subtraction between c++ and python

[Jing Ye]

What's the difference between the following c++ code and python code? Both of them will load a same test.jpg(32x32) and subtract a mean matrix (255, 0, 0).

But the output images are apparently different. it's very strange.

Can any one help me?

test.jpg is

#include <caffe/caffe.hpp>

#include <opencv2/core/core.hpp>

#include <opencv2/highgui/highgui.hpp>

#include <opencv2/imgproc/imgproc.hpp>

#include <algorithm>

#include <iosfwd>

#include <memory>

#include <string>

#include <utility>

#include <vector>

using namespace caffe;

using std::string;

int main(int argc, char** argv) {

    cv::Mat img = cv::imread("test.jpg", CV_LOAD_IMAGE_COLOR);

    cv::Mat mean = cv::Mat(32, 32, CV_32FC3, cv::Scalar(255, 0, 0));

    cv::Mat img_float;

    img.convertTo(img_float, CV_32FC3);

    cv::Mat img_normalized;

    cv::subtract(img_float, mean, img_normalized);

    // output

    cv::imwrite("1.jpg", img_normalized);

}

c++ output: 1.jpg is

----------------------------------------------------------------------------------------------------------------------------

import numpy as np

import caffe

import scipy.misc

from PIL import Image

transformer = caffe.io.Transformer({'data': (1, 3, 32, 32)})

transformer.set_transpose('data', (2,0,1))  # move image channels to outermost dimension

transformer.set_mean('data', np.array([255, 0, 0])) # subtract the dataset-mean value in each channel

transformer.set_raw_scale('data', 255)      # rescale from [0, 1] to [0, 255]

transformer.set_channel_swap('data', (2,1,0))  # swap channels from RGB to BGR

image = caffe.io.load_image('test.jpg')

transformed_image = transformer.preprocess('data', image)

# save result

t = transformed_image.transpose((1,2,0))

scipy.misc.imsave('2.jpg', t)

python output: 2.jpg is