Fix mAP calculation for areas with 0 predictions

[23pointsNorth]

The issue was first dicussed here: #1061 (comment)

In cases of iou_type segm mAP calculation with 0 predictions, the standard box_area calculation is needed. This is fixed to use the updated compute_area method.

Sidenotes:

• if gt_label_mask is with len>0, accessing a python list would fail -> need to casting
• compute_area assumes a list of areas to be computed, thus the need to expand dims in cases a single element is selected from the array.

Has been tested to work with bbox and segm.

What does this PR do?

Fixes part of a bug first highlighted as in #1061.

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

Would appreciate if this is added to the next patch release! Maybe I was too late for 0.9.1 :(

[Borda]

Can you add a test for this configuration? 🐰

[codecov]

Codecov Report

Merging #1080 (5a54a39) into master (9962764) will decrease coverage by 0%.
The diff coverage is 100%.

@@          Coverage Diff           @@
##           master   #1080   +/-   ##
======================================
- Coverage      95%     95%   -0%     
======================================
  Files         181     181           
  Lines        8103    8103           
======================================
- Hits         7672    7660   -12     
- Misses        431     443   +12     

[23pointsNorth]

Hi @Borda ,

I am having issues running the test locally (make test assumes to have metrics/_data/detection/instance_segmentation_inputs.json, and those I assume are fetched separately not through e.g. make download-data).

I've added what I expect to be a sensible test (0 gt, 0 predictions), but hard to evaluate if it's the right spot or if it impacts some of the numerical test you have already set.

Or laternatively make a case like this with flipped gt/pred?

[Borda]

I am having issues running the test locally (make test assumes to have metrics/_data/detection/instance_segmentation_inputs.json, and those I assume are fetched separately not through e.g. make download-data).

good point we will add it to our make file...
the download link is https://pl-public-data.s3.amazonaws.com/metrics/data.zip

[23pointsNorth]

Thanks! I assume also in .requirements/text.txt one needs to add tensorflow (because of bert -> huggingface_hub -> keras_mixin -> tf)

ERROR tests/text/test_bertscore.py - ModuleNotFoundError: No module named 'tensorflow'

Is there a way to run just detection tests?

edit: python3 -m pytest tests/detection should do it.
edit2: After the latest push, the above command makes the tests/detection/mean_ap.py finish.

[23pointsNorth]

Do you folks have any pointers to when the next patch version would be released?

[Borda]

Do you folks have any pointers to when the next patch version would be released?

we can do one next week :)

[dcela]

Ran into the same issue, for now will build from current master!

[dcela]

If there is interest in getting people just to improve documentation, I would be willing to help!

[Borda]

If there is interest in getting people just to improve documentation, I would be willing to help!

please do 🐰