CoreML performance cost of Vision Framework vs. CVPixelBuffer?

Question

There are two options for passing images into CoreML models:

1. Pass in a CGImage to Vision Framework with its niceties
2. Pass in a CVPixelBuffer directly

Is there any data on the memory and processing overhead associated with using the Vision framework vs. passing a CVPixelBuffer directly to CoreML?

Thoughts based on what I've seen while debugging:

Memory

Assuming we already have the data in a CVPixelBuffer, creating the CGImage to pass to Vision seems to double the memory usage. It looks like Vision is creating a new object in CoreVideo/CoreImage from createPixelBufferFromVNImageBuffer which makes sense as it needs to create a copy of the image to crop/rotate/scale.

Processing

You're going to have to do the rotation and/or scaling either way. I'd assume Vision does those at least as efficiently as you could do by hand with Accelerate. So should not be any overhead here.