Introduction to Two-Stage Object Detection Architectures

Two-stage architectures are the premises for this special task and bring a huge novelty and progress against the traditional object detection methods. We can appreciate their methodologies to bring a novel approach to object detection. However, they are slow and a bit outdated compared to later discovered one-stage methodologies and their improved versions.

Whether to use it or not, knowing the logic behind two-stage architectures is important. On the other hand, if model speed is not our priority (if we don’t work on a real-time project, for example), we can still use the improved versions of two-staged detection models. Similarly, understanding the methodology of individually extracting the region of interest is important since this approach has a wide range of uses to integrate into other networks for different tasks.

Models in two-stage object detection

Region-based convolutional neural networks are the most popular two-stage object detection models. We will review the region-based convolutional neural network family from older to newer versions:

• R-CNN (Region-based convolutional neural networks)
• Fast R-CNN
• Faster R-CNN