What is the core architectural difference between one-stage and two-stage detectors?
Two-stage detectors split the work in two: a first network proposes candidate regions, then a second classifies and refines each one. One-stage detectors skip the proposal step and predict all boxes and classes in a single forward pass. Two-stage tends to be more accurate on small or dense objects; one-stage is far faster and is what you use for real-time video.
Two-stage (e.g. Faster R-CNN)
Stage one is a Region Proposal Network that says 'there's probably something here' for a few thousand candidate regions. Stage two crops features for each proposal and classifies it, refining the box. The extra stage costs time but lets the model look carefully at each candidate, which historically won on accuracy.
One-stage (e.g. YOLO, SSD)
The network divides the image into a grid and, in one pass, predicts boxes and class probabilities directly at every location. No separate proposal step. That single-pass design is why YOLO can run at high FPS on video.
The old rule was 'two-stage for accuracy, one-stage for speed', but modern one-stage detectors have largely closed the accuracy gap while keeping their speed edge, which is why nearly all real-time systems today are one-stage.
Related questions
Hands-on tutorials
Stuck on this in a real project?
Book a free call and I’ll give you a straight answer on your specific case.
Book a consultation