Region Proposals

Instead of using a CNN to detect objects at every location in an image, "region proposals" works by first determining areas of interest (i.e. areas which are more likely to contain an object) using segmentation, and then runs the CNN only on those regions, making the process more computationally efficient.

This is also referred to as R-CNN (Regions with CNN).

This implementation of R-CNN was relatively slow because it processed the selected regions one at a time.

"Fast R-CNN" was later proposed and it used a convolutional implementation of sliding windows (as discussed earlier), so as to process all the chosen regions simultaneously.

"Faster R-CNN" used a CNN instead of traditional segmentation techniques to propose regions, making the overall process even faster.