region proposal

A region proposal is a candidate box that says 'an object might be here, worth a closer look' — without yet saying what the object is. The idea is to split detection into two cheaper questions. Rather than running a full classifier at every possible location, scale, and aspect ratio (millions of windows), you first generate a few hundred or few thousand promising regions, then spend the expensive classification only on those. Proposals dramatically shrink the search space.

Early proposal methods were classical, image-based algorithms that ran before any deep network. Selective Search (used by R-CNN and Fast R-CNN) over-segments the image into superpixels by color and texture, then greedily merges similar neighbors into a hierarchy of regions, emitting the merged boxes as proposals — roughly 2,000 per image. EdgeBoxes scored boxes by how many contours are wholly enclosed. These are class-agnostic: they aim for high recall of 'objectness', trusting the later classifier to label and filter.

The key metric for a proposal set is recall at a given proposal count: what fraction of true objects has at least one proposal overlapping it (IoU above threshold), using only N proposals. A good generator achieves high recall with few proposals, because every extra proposal costs downstream compute. There is a direct speed/accuracy trade: more proposals catch more objects but slow the detector.

The watershed change was making the proposal step learnable and shared with the detector backbone. Faster R-CNN replaced Selective Search with a region proposal network (RPN) — a small convolutional head on the shared feature map — turning proposals from a fixed external preprocessing step into a trained, GPU-friendly part of the model. This is the conceptual hinge between first-generation detectors and modern two-stage ones.