core: inconsistency of addition for multi-channel Mat and Scalar

[alalek]

OpenCV 3.4.9 / 4.2.0:

arg1 - uint8 matrix with 3 channels:

• Mat m__1_2_3(Size(2, 2), CV_8UC3, Scalar(1, 2, 3));

arg2:

• 5
• Scalar(5) (identical to Scalar(5, 0, 0, 0))
• Scalar(5, 1e-6, 0)
• Scalar::all(5) (identical to Scalar(5, 5, 5, 5))

Result variants:

1. [6, 2, 3] - modified 1st channel only
2. [6, 7, 8] - all channels are modified

#	C++ code	arg1 5	arg2 Scalar(5)	arg3 Scalar (5, 1e-6, 0)	arg4 Scalar::all(5)
1	cv::add(arg1, arg2, res)	2	1	1	2
2	MatExpr res = arg1 + arg2	1 #14738	1	1	2
3	MatExpr res = 1.001 * arg1 + arg2	2 #14738	2 #16538	1	2

Analyse:

• results with Scalar(5, 1e-6, 0)(arg3) and Scalar::all(5)(arg4) are consistent and adequate
• some other cases should be be deprecated / changed / removed?
• case with numeric scalar (5) may have designed to use with complex numbers (emulation through matrices with 2 channels) where 5 is interpreted as a real number.
• current implicit conversion of 5 into Scalar (no explicit bug/feature) converts this into Scalar(5, 0, 0, 0). Extra overloads may help to avoid implicit conversions (but may break user code).

Related unclear stuff:

• multi-channel Mat::eye() without element initializer.
• missing explicit for Vec and Scalar constructors: cv::Matx33f tmp = true compiles #11956
• almost zero tests for MatExpr stuff

[saskatchewancatch]

Looking at the Jiras, it seems many people expect OpenCV matrix's to behave just like in Matlab.
Perhaps behaviour should strive to be as close as possible to the way Matlab treats Matrix's?

[zchrissirhcz]

I'm just testing a very simple code, comparing Python OpenCV & C++ OpenCV's results. They ain't the same.

System Version?

Windows 64 Bit

OpenCV Version?

OpenCV 4.3.0 Windows prebuilt package, for both C++ & Python interface.

Python Version?

E:\dbg\zz\test>python
Python 3.7.4 (default, Aug  9 2019, 18:34:13) [MSC v.1915 64 bit (AMD64)] :: Anaconda, Inc. on win32

PATH?

`D:\lib\opencv\4.3.0\build\x64\vc15\bin;` is in the very begging of PATH

Reproduceble Code?

# coding: utf-8
import numpy as np

import sys
sys.path.insert(1, 'D:/lib/opencv/4.3.0-prebuilt/python/cv2/python-3.7')

import cv2

im = np.ones((100,100,3))
im = im * 0.73 * 255
#print(im)
cv2.imwrite('res-py.bmp', im)

#include <opencv2/opencv.hpp>

int main(){
    cv::Mat im = cv::Mat::ones(100, 100, CV_64FC3);
    im = im * 0.73 * 255;
    cv::imwrite("res-cpp.bmp", im);
    return 0;
}

Result?

The pasted Python snippet generates a "gray color" image, while the pasted C++ snippet generates a "blue color" image.

Expected result?

The C++ code should generate "gray color" image instead of "blue image" (only the first channel is modified)

[alalek]

Numpy doesn't have "channels" like OpenCV.

This problem is not related to MatExpr - difference is in array creation itself. Use this instead:

cv::Mat im(100, 100, CV_64FC3, Scalar::all(1));

[zchrissirhcz]

@alalek OK, figured out by reading mat.hpp's comment of static MatExpr ones(int ndims, const int* sz, int type);:

@note In case of multi-channels type, only the first channel will be initialized with 1's, the
others will be set to 0's.

This reminds me that MathExpr is more similar with Matlab grammar instead of Python. LOL.