Fix overlapping masks in semantic segmentation targets

[sokovninn]

This PR resolves the issue of overlapping masks in semantic segmentation targets.

Benchmark Results (1 epoch COCO, 5000 images, 8 workers):

• Before: 4:00 minutes
• After: 4:13 minutes (approximately 6% increase in processing time)

Notes:

The added overhead is due to the additional processing required to remove overlaps in the segmentation masks. Further optimizations may be explored if necessary.

Update:

After further optimizations there is no overhead anymore.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.34%. Comparing base (6bcbb11) to head (b8e500a).
Report is 26 commits behind head on main.

✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #215      +/-   ##
==========================================
+ Coverage   92.33%   92.34%   +0.01%     
==========================================
  Files          79       80       +1     
  Lines        4161     4324     +163     
==========================================
+ Hits         3842     3993     +151     
- Misses        319      331      +12     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[kozlov721]

(approximately 6% increase in processing time)

IMHO the overhead is too big. We should check and correct this on the side of LuxonisDataset when the dataset is being created.

[kozlov721]

Thanks, LGTM! Could you just add a small test for this?