[ad_1]
I’m engaged on an app with a segmentation portion. We have now tried passing the next sizes to VNGeneratePersonSegmentationRequest:
- 10×10
- 100×100
- 256×256
- 512×512
- 4096×4096
Even after the primary (Warmup) request, we’re discovering efficiency among the many first 4 classes to be virtually the identical (~60 ms common on iPhone 13), and in some way 4096×4096 is simply 2(!) instances slower than 512×512.
That is really a tremendous outcome if you’re searching for very excessive decision segmentation information, sadly, we are attempting to see if we will nonetheless generate QualityLevel.correct masks however save on the efficiency hit by processing at a decrease decision.
Has anybody had any success optimizing the VNGeneratePersonSegmentationRequest to run on decrease decision inputs? Probably I am lacking one thing by way of pixel buffer format (utilizing 32BGRA), or enter kind? Assuming this makes use of the GPU, is there any approach to seize some extra of the internals of the neural networks/shaders powering the precise segmentation algorithm?
Thanks,
Dennis
[ad_2]
