Feature/make sv.mask annotator as fast as annotators in popular detection frameworks #472

[xenteros]

Description

• Combined all colored masks before applying them to the scene
• Made sure that returned ndarray is np.uint8 for video writer simplicity

#472

• Bug fix (non-breaking change which fixes an issue)

How has this change been tested, please provide a testcase or example of how you tested the change?

I profiled the code with VizTracer and compared outputs frame by frame and they are equal.

[CLAassistant]

All committers have signed the CLA.

[SkalskiP]

Hi, @xenteros! 👋🏻 Thanks for your interest in Supervision. I'm benchmarking your solution right now.

[SkalskiP]

Hi @xenteros 👋🏻, I've been busy making tests of this change. Here is what I got:

1. Consistency of obtained results

• develop

• xenteros:feature/Make_sv.MaskAnnotator_as_fast_as_annotators_in_popular_detection_frameworks_#472

2. Drawing speed

Annotation speed depends on external factors - other processes running on the machine and varies between test runs. To validate the results, I performed 5 tests with the code in the current develop and 5 with the code from xenteros:feature/Make_sv.MaskAnnotator_as_fast_as_annotators_in_popular_detection_frameworks_#472. I averaged the results.

• develop

  reference (only inference)	inference + box annotation	inference + mask annotation
  123 sec	125 sec	298 sec
  125 sec	125 sec	300 sec
  124 sec	125 sec	301 sec
  124 sec	125 sec	299 sec
  124 sec	125 sec	301 sec
  124 sec (average)	125 sec (average)	300 sec (average)

  Annotation overhead: 300 - 124 = 176
  Per frame annotation overhead: 350 / 341 = 0.51
  Google Colab - link

• xenteros:feature/Make_sv.MaskAnnotator_as_fast_as_annotators_in_popular_detection_frameworks_#472

  reference (only inference)	inference + box annotation	inference + mask annotation
  125 sec	125 sec	159 sec
  124 sec	125 sec	159 sec
  124 sec	125 sec	159 sec
  124 sec	125 sec	162 sec
  124 sec	126 sec	161 sec
  124 sec (average)	125 sec (average)	160 sec (average)

  Annotation overhead: 160 - 124 = 36
  Per frame annotation overhead: 36 / 341 = 0.11
  Google Colab - link

3. Summary:

• Results are different - in the case of overlapping masks, we see only the one that was applied to the area last
• Updated MaskAnnotator is 5x faster

[SkalskiP]

Speed comparison across all three MaskAnnotator versions

[SkalskiP]

LGTM. Merging! @xenteros, thanks a lot for your contribution. This annotator was a big performance bottleneck.