Faster R-CNN
Faster Region-based CNN (R-CNN) is a significant advancement over its predecessors, R-CNN and Fast R-CNN. It incorporates region proposal generation within the network architecture, eliminating the need for external algorithms. In traditional approaches, region proposals were generated using selective search, which could be time-consuming. Fast R-CNN improved upon this by performing a single forward pass of the CNN on the entire input image, generating region proposals based on the extracted features. However, Faster R-CNN further enhanced the architecture by introducing a Region Proposal Network (RPN). The RPN shares convolutional layers with the object detection network and generates region proposals by sliding a small network window over the feature map. Simultaneously, it predicts objectness scores and refines bounding box coordinates. The Faster R-CNN architecture enables end-to-end training and faster inference by efficiently sharing convolutional layers between the RPN and the object detection network.