[FIX] Dataset Adapter: Avoid duplicated annotation and permit empty image

[sungmanc]

This PR includes

• VOC import error: VOC includes several annotation formats (label, bbox, mask). In this case, if we want to make detection dataset, then we need to filter out the mask, and label annotations. --> len(shapes) >0.
• However, if there are empty label image that has no annotations (len(annotations)=0), we need to add them to DatasetItemEntity. --> len(datumaro_item.annotations)==0.
• For the unlabeled dataset, it also be included in DatasetItemEntity. --> subset == Subset.UNLABELED

if len(shapes) > 0 or subset == Subset.UNLABELED or len(datumaro_item.annotations) == 0:

---

• To test the empty_label case, I modified the test data that includes the empty_label image.

• Implement unit-test to validate whether empty_label image is successfully included in DatasetItemEntity or not.

  

• Test results
  

Thanks for reporting this issue. @eunwoosh , @supersoob

[codecov-commenter]

Codecov Report

Patch coverage: 100.00% and project coverage change: -0.01 ⚠️

Comparison is base (0e4a8b7) 80.54% compared to head (044e2c1) 80.54%.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1873      +/-   ##
===========================================
- Coverage    80.54%   80.54%   -0.01%     
===========================================
  Files          484      484              
  Lines        33005    33007       +2     
===========================================
  Hits         26585    26585              
- Misses        6420     6422       +2     

Impacted Files	Coverage Δ
.../core/data/adapter/segmentation_dataset_adapter.py	94.01% <ø> (ø)
otx/core/data/adapter/detection_dataset_adapter.py	100.00% <100.00%> (ø)

... and 3 files with indirect coverage changes

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[eunwoosh]

I think dataset w/o annotations need to be filtered out during training stage. I run detection on this branch but it isn't.

[sungmanc]

@eunwoosh , I added logic to avoid inputting the empty image to train dataset. And changed tests/assets.

[eunwoosh]

LGTM.

[sungmanc]

@supersoob , @jaegukhyun could you recheck this?