Two-Stage vs One-Stage Detectors
What You'll Learn
You'll understand the two fundamental architectural paradigms for object detection: the slower but precise two-stage approach versus the faster one-stage direct detection strategy.
The Core Architectural Split
Object detection networks solve both localization (where is it?) and classification (what is it?). How they solve these problems divides detectors into two families:
Two-Stage Detectors
These networks work in two distinct phases:
- Region proposal stage: Generate candidate bounding boxes (regions of interest) that might contain objects
- Classification stage: For each proposal, classify the object and refine the bounding box
Think of this like a two-person team: one person points out "there might be something here," then a specialist examines each suggestion carefully.
Examples: R-CNN family (you'll learn these later)
One-Stage Detectors
These networks perform detection in a single pass. They simultaneously predict class probabilities and bounding box coordinates directly from the image, without separate region proposals.
Think of this as one expert who looks at the entire scene once and immediately tells you what's where—faster but potentially less careful about tricky cases.
Examples: YOLO, SSD (coming soon in your learning path)
The Trade-off
Two-stage detectors: Higher accuracy, especially on small or overlapping objects, but slower inference time
One-stage detectors: Real-time performance, simpler architecture, but historically slightly lower accuracy (though the gap has narrowed)
The choice depends on your application: autonomous vehicles need speed, medical imaging prioritizes precision.
Key Takeaway: Two-stage detectors separate region proposal and classification for higher accuracy, while one-stage detectors predict everything directly for speed—a fundamental accuracy-speed trade-off in detection architectures.