How do two-stage detectors differ from single-stage detectors in architecture, speed, and quality?

Two-stage detectors (e.g., Faster R-CNN) and single-stage detectors (e.g., YOLO, SSD) differ in architecture, speed, and quality as follows:

• Architecture:

  • Two-stage detectors first generate region proposals, then classify and refine these regions. This is a two-step process.
  • Single-stage detectors directly predict classes and object coordinates at all grid positions without a separate proposal generation step.

• Speed:

  • Single-stage detectors are usually faster as they make all predictions in one pass.
  • Two-stage detectors are slower due to the additional proposal processing step.

• Quality (accuracy):

  • Two-stage detectors generally achieve higher accuracy, especially on complex tasks and small objects.
  • Single-stage detectors may be less accurate, but modern models have significantly reduced this gap.

Example: Faster R-CNN first generates region proposals, then classifies them, while YOLO directly predicts classes and coordinates for each grid cell.