In machine learning, Region of Interest (ROI) proposal is a method for subsampling an image to identify "interesting" subregions. A canonical example is from Region Convolutional Neural Networks, where the task is to identify objects in a scene:
The point of the above ROI proposal is that it makes subsequent multiple object detection more efficient since the entire image no longer needs to be processed and because most objects will fall within the detected subregions. To be clear, ROI proposal is usually done using lower-level methods like looking at color gradations and estimating region boundaries. Nowedays, ROI proposal is done with convolutional layers, but even these use fairly low-level kernels to quickly propose regions. I'm specifically interested in the convolutional approaches, as they seem most relevant to the brain.
I'm wondering if there is any evidence of such ROI proposal in animal vision? Specifically what is our current state-of-the-art understanding of how the brain identifies multiple objects in a complex scene? Do we have reason to believe that it's not just a left-right and top-down scan of the image, but something more efficient where lower-level neurons propose subregions to explore?
It's been difficult to search for this topic online because it mainly links to ROI proposal for brain scans.
