Currently, there is no proper support for handling of empty matrices or distinguish them from just declared cv::Mat m;.

AFAIK, meta information should not be used if Mat is .empty().

See also: #7710

I understand we cannot differentiate with a proper empty matrix created by cv::Mat m (which has a default type of CV_8UC1).
Still, behavior is inconsistent: resize(0) keeps the type (cf

My patch is actually to solve: cv::Mat(0,0,CV_16SC3).clone() returns cv::Mat(0,0,CV_8UC1) (and not an empty matrix of the right type CV_16SC3).

So please advise on how we should fix the following (maybe by just updating the doc in the right places):

• resize(0) keeps the type
• clone() does not keep the type for an empty matrix (which contradicts the documentation at

  opencv/modules/core/include/opencv2/core/mat.hpp

  Line 1195 in b19697e

  The method creates a full copy of the array. The original step[] is not taken into account. So, the

  )
• copyTo() does not transmit the type for an empty matrix because it does not call .create (which contradicts the documentation at

  opencv/modules/core/include/opencv2/core/mat.hpp

  Line 1204 in b19697e

  m.create(this->size(), this->type());

  )
• release() removes the type if deallocated (and this is not documented at

  opencv/modules/core/include/opencv2/core/mat.hpp

  Line 1495 in b19697e

  /** @brief Decrements the reference counter and deallocates the matrix if needed.

  )

I still think my patch would bring some consistency but I would also need to have release() keep the type (which I can add).

release()

Main point here that "Mat" is a kind of smart pointer over some buffer.
To behave similar to smart pointers, .release() should reset instance to the state of empty and undefined cv::Mat m;.

which contradicts the documentation

Currently it is a gray zone which is not documented.

I agree with .release(), the semantic needs to be clarified in the doc though.

There is actually a way to differentiate between cv::Mat() and cv::Mat(0,0,CV_8U) though their type is 0 for both: elemSize() is 0 in the first case, 1 in the other case.

Probably, we should use the check, proposed by @vrabaud (e.g. m.total() == 0 && m.elemSize() == 0) to check for empty matrices and produce meaningful results in this case instead of crashing

instead of crashing

where is code crashing?

m.elemSize() == 0

Even cv::Mat m; returns 1 here (as its type is CV_8UC1 = 0), so this check doesn't make a lot of sense.

.elemSize() does return different results for me with 4.5.3.

#include "opencv2/core.hpp"
#include <iostream>


int main(int, char**argv){
  cv::Mat m1, m2(0, 0, CV_8U);
  std::cout << m1.elemSize() << " " << m2.elemSize() << std::endl;
}

And

clang++ -l opencv_core -I /usr/include/opencv4/ -o test main.cc
./test
0 1

Maybe we need a .empty() like we have right now that just checks if there are no elements, and a .uninitialized() / .released() / emptyFull() that checks whether we have a cv::Mat() (no element, no type).

.elemSize()

Right, it has dedicated handling for empty Mat:

I messed that with CV_ELEM_SIZE(flags).

Also please note CV_DbgAssert() check in the code above - using .elemSize() with empty Mat triggers error.

However original expression form m.total() == 0 && m.elemSize() == 0 doesn't make sense.

Hi, what do I do with that PR ? Port it to 4.x or give up until 5.x which has better handling of 0d ?