Slicing to downsample image does not work

[sfe-SparkFro]

Using a 640x480 image (source: Wikipedia), I attempted to downsample it for display with cv2.imshow(display, img[::2, ::2]). In standard Python, the following was displayed:

In MicroPython, the following was displayed:

Poking around with a debugger right here:

micropython-opencv/src/convert.cpp

Line 113 in da558f4

Mat mat = Mat(ndims, size, type, (void*) ndarray->array, step);

The constructed Mat object's step array does not match the provided step array:

(gdb) p step
$8 = {3840, 6, 0 <repeats 31 times>}
(gdb) p mat.step.buf
$9 = {3840, 3}

I suspect the problem is here:

@param steps Array of ndims-1 steps in case of a multi-dimensional array (the last step is always set to the element size)...

Not sure why there's a discrepancy between standard Python and MicroPython, because ndarray_to_mat() is basically the same as pyopencv_to() (at least, as far as how step is constructed). I suspect when standard NumPy does the array slicing, it makes a copy instead of reusing the same array like ulab NumPy does, so forcing the last element of steps to be the element size isn't a problem. Maybe. Regardless, might just be able to force the Mat object's step array to match.

[sfe-SparkFro]

Simple adding mat.step.buf[ndims-1] = step[ndims-1]; after here:

micropython-opencv/src/convert.cpp

Line 113 in da558f4

Mat mat = Mat(ndims, size, type, (void*) ndarray->array, step);

does not resolve the problem.

[sfe-SparkFro]

I suspect the problem is here:

@param steps Array of ndims-1 steps in case of a multi-dimensional array (the last step is always set to the element size)...

That indeed appeared to be the root problem.

Not sure why there's a discrepancy between standard Python and MicroPython, because ndarray_to_mat() is basically the same as pyopencv_to() (at least, as far as how step is constructed). I suspect when standard NumPy does the array slicing, it makes a copy instead of reusing the same array like ulab NumPy does

step was not the problem, it was _strides, which is used to set step. _strides gets changed if needcopy == true (which it is if _strides[ndims-1] != elemsize).

Adding the copy if needed in 6a81d86 fixed the problem.