Add mean average precision metric for object detection

[tkupek]

Mean Average Precision (mAP) for object detection

New metric for object detection.

What does this PR do?

This PR introduces the commonly used mean average precision metric for object detection.
As there are multiple different implementations, and even different calculations, the new metric wraps the pycocotools evaluation, which is used as a standard for several academic and open-source projects for evaluation.

This metric is actively discussed in issue, resolves #53

TODO

• check if pycocoeval can handle tensors to avoid .cpu() calls (it cannot)
• standardize MAPMetricResults to have all evaluation results in there
• refactor some code parts (e.g. join get_coco_target and get_coco_preds methods)
• add unittests and documentation in torchmetrics format

Note

This is my first contribution to the PyTorchLightning project. Please review the code carefully and give me hints on how to improve and match your guidelines.

[codecov]

Codecov Report

Merging #467 (13ec1ee) into master (b2240e1) will decrease coverage by 0%.
The diff coverage is 93%.

@@          Coverage Diff           @@
##           master   #467    +/-   ##
======================================
- Coverage      95%    95%    -0%     
======================================
  Files         148    150     +2     
  Lines        5142   5307   +165     
======================================
+ Hits         4897   5050   +153     
- Misses        245    257    +12     

[Borda]

Any why is this massive test failer for PT 1.4?

E   ValueError: numpy.ndarray size changed, may indicate binary incompatibility. Expected 88 from C header, got 80 from PyObject

Update: https://stackoverflow.com/questions/66060487/valueerror-numpy-ndarray-size-changed-may-indicate-binary-incompatibility-exp

[senarvi]

This PR looks great to me.

[tkupek]

Thank you everybody for your support 🙂

[Borda]

Thank you everybody for your support slightly_smiling_face

the pleasure is on our side, Thx for your work and patience... 🐰

[SkafteNicki]

Thanks everybody for their input on getting this metric implemented 😄
If anybody feels like it, it would be great to have an example in this folder:
https://github.com/PyTorchLightning/metrics/tree/master/tm_examples
as this is a bit more complicated metric to use than other.

[elstonaug]

Hello, One question - do we need to exclude class 0 which is the background class when updating / handling the metrics?