R-CNN
/ AR-C-N-N /
R-CNN is a family of object-detection methods that solved a 2013-era puzzle: how do you find objects when you don't know how many there are or where? Its answer was a two-step strategy. First, propose a few thousand candidate regions — boxes that might contain something. Then run an image classifier on each box to decide what (if anything) is inside, and refine the box's edges. "R" is for region; the CNN is the convolutional neural network that does the recognizing.
The original R-CNN worked but was painfully slow, because it cropped and re-analyzed every one of those thousands of regions separately. Two follow-ups fixed this. Fast R-CNN analyzed the whole image once and then looked up each region in that shared analysis, instead of redoing it. Faster R-CNN went further and let a small neural network propose the regions too, so the entire pipeline became one trainable system. Each step traded clever hand-engineering for learned, faster components.
Why remember R-CNN? It established the two-stage template — propose, then classify — that defined accurate detection for years and still underlies many high-precision systems. Its trade-off is the headline: two-stage detectors tend to be more accurate but slower than the one-pass detectors (like YOLO) that came after. For tasks where missing an object is costly and you can spare the compute, the R-CNN lineage is often still the careful choice.
On a kitchen photo, Faster R-CNN first proposes ~300 candidate boxes from a quick scan. It then classifies each: most are rejected as background, but a handful survive as "mug 0.91," "toaster 0.87," "banana 0.66," with their box edges nudged to hug each object.
The two-stage idea in one frame: propose many regions, then classify and refine the survivors.
"R-CNN" usually refers to the whole lineage — R-CNN, Fast R-CNN, Faster R-CNN — not one fixed model. The common thread is the two-stage "propose then classify" design, which trades speed for accuracy compared with single-stage detectors.