MultiTemplateMatching Python Versions Save

Changed licence to MIT for easier integration with other projects

Minor version that throws a more explicit error message when providing 0-width or height image or template. Also added pip install matplotlib to tutorial notebooks since not an dependency of MTM.

Provide faster search with mulitple templates by parallelizing the search -> PR by @bartleboeuf (thanks !)

Update internal NMS code to work with latest OpenCV, and force opencv version to min 4.5.4, Fix #14

The version fixes issue with the latest version of opencv (4.5.4, maybe before)), for which cv2.NMSBoxes return an array of different shape. Just by limiting the version of OpenCV to max 4.3.0.
Next release of MTM will be modified to use the latest version of OpenCV.

As suggested in issue #11, the size of the template is checked before any computation to make sure they fit in the search region or image.

Also updated requirements and packaging (pyproject.toml)

The comparison of the pixel values can be limited to a fraction of the template by providing an associated mask image, which should have the same dimension than the template. To do so the tuple for a given template should be ("label", template_array, mask_array)

using a mask is only possible with method 0/TM_SQ_DIFF (not compatible with the Non-Maxima suppression though) and 3/TM_CCORR_NORMED

using a mask is optional (and not always beneficial) the previous format ("label", template_array) is still acceptable and within the list of templates, some can have a mask while other not.

See new tutorial with mask.

See release note for 1.5.3.post2

• Now findMatches return an empty dataFrame with proper column names if no object is found (prevent an issue with NMS then)
• better handle the different template matching method (method 0/SQDIFF is supported by some functions, but not matchTemplates)

This version replaces the original hard-coded NMS with the opencv cv2.dnn.NMSBoxes method.
This simplifies the code and improve portability to other programing languages.
Performance-wise it is similar to potentially slightly faster.
The function signature is the same except the default scoreThreshold is now 0.5

Other changes:

• the method in MTM.NMS PointInRect and computeIoU have been removed (not used anymore for NMS)
• tutorials updated with timeit benchmarks