faster sv.MaskAnnotator

[kapter]

Description

This PR is related to #417
I've made some enhancements to sv.MaskAnnotator.

1. calc bgr color once for color mask creation
2. replace np.where with cv2.add operation for scene with hole and colored object

Result

speed up two times

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Hello there, thank you for opening an PR ! 🙏🏻 The team was notified and they will get back to you asap.

[SkalskiP]

Please remove that extra argument.

[SkalskiP]

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

1. Consistency of obtained results

• develop

  

• kapter:Make-MaskAnnotator-faster

  

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 4 tests with the code in the current develop and 4 with the code from kapter:Make-MaskAnnotator-faster. I averaged the results.

• develop

  baseline (only inference)	inference + box annotation	inference + mask annotation
  108 sec	127 sec	451 sec
  103 sec	108 sec	465 sec
  129 sec	114 sec	475 sec
  128 sec	127 sec	476 sec
  117 sec (average)	119 sec (average)	467 sec (average)

  Annotation overhead: 467 - 117 = 350
  Per frame annotation overhead: 350 / 341 = 1.03
  Google Colab - link

• kapter:Make-MaskAnnotator-faster

  baseline (only inference)	inference + box annotation	inference + mask annotation
  104 sec	103 sec	279 sec
  99 sec	98 sec	296 sec
  96 sec	104 sec	281 sec
  102 sec	104 sec	266 sec
  100 sec (average)	102 sec (average)	280 sec (average)

  Annotation overhead: 280 - 100 = 180
  Per frame annotation overhead: 180 / 341 = 0.52
  Google Colab - link

3. Summary:

• Results are consistent - no visual difference
• Updated MaskAnnotator is 0.5 s per frame faster

[SkalskiP]

@kapter I'm curious if we can make it even faster. 🤔 💭 0.5 sec per frame is 2x faster than the current implementation, but it is still 0.5 sec. It is far from real-time annotation.

Having said that, I'm merging this PR. It is a clear improvement.

Maybe we could somehow sacrifice the quality of annotation to get a speed boost.

[kapter]

I missed
Check this please https://github.com/kapter/supervision/tree/MaskAnnotator_faster_again