Vision · 2014–2017 · two-stage detection
R-CNN Family
Region-based detectors: propose candidate objects, then classify and refine each proposal. Faster R-CNN made proposal generation part of the network itself.
ModLensVision
Interactive Diagram
Focus the lens
Click any component to read what it does. Signal direction follows the edges.
inputImage. Scene enters the shared backbone.
Core idea
Attend before you decide: a region proposal network tells the classifier where to look, trading a little speed for precise, high-recall detection.
Why it exists
Sliding-window classification is wasteful; proposals focus compute on plausible objects and improve localization.
Data Flow
What moves through the system
- 01Backbone extracts features for the whole image.
- 02Region Proposal Network suggests candidate boxes.
- 03RoI pooling reads features for each proposal.
- 04Heads classify proposals and refine box coordinates.
Strengths
- + High accuracy and localization quality
- + Conceptually clean two-stage design
Limitations
- − Slower than one-stage detectors
- − More moving parts to train and tune
Applications
- · Precision inspection
- · Satellite imagery
- · Where accuracy outweighs latency