The first step in the Region-based Convolutional Neural Network (R-CNN) pipeline is to perform selective search to extract multiple high-quality region proposals from the input image. These regions are typically selected at various scales, shapes, and sizes. Each of these region proposals is then labeled with a ground-truth class and a bounding box to be used in subsequent stages.

Selective Search in R-CNN

Although the Region-based Convolutional Neural Network (R-CNN) model effectively utilizes pretrained Convolutional Neural Networks (CNNs) to extract features, it suffers from significant computational inefficiency. Because the model relies on extracting thousands of region proposals from a single input image, it must execute thousands of separate CNN forward propagations to perform object detection. This massive computational burden makes the standard R-CNN architecture too slow and infeasible for widespread use in real-world applications.

Computational Bottleneck of R-CNN

In object detection, a region proposal is a candidate bounding box extracted from an input image that potentially contains an object of interest. Models typically extract a large number of these proposed regions (e.g., $$2000$$ per image) to process further. Additionally, predefined anchor boxes can also serve as region proposals.

Region Proposal

The Region-based Convolutional Neural Network (R-CNN) is an object detection model that operates by first extracting a large number of region proposals (e.g., $$2000$$) from an input image. Anchor boxes can also be considered as such proposals. A Convolutional Neural Network (CNN) is then used to perform forward propagation on each individual region proposal to extract its features. Finally, these features are utilized to predict the class and bounding box for each proposed region.

Claude

There are a few different algorithms for object detection and they can be split into two groups:

1. Algorithms based on classification. They are implemented in two stages. First, they select regions of interest in an image. Second, they classify these regions using convolutional neural networks. This solution can be slow because we have to run predictions for every selected region. A widely known example of this type of algorithm is the Region-based convolutional neural network (RCNN) and its cousins Fast-RCNN, Faster-RCNN and the latest addition to the family: Mask-RCNN. Another example is RetinaNet.

2. Algorithms based on regression – instead of selecting interesting parts of an image, they predict classes and bounding boxes for the whole image in one run of the algorithm. The two best known examples from this group are the YOLO (You Only Look Once) family algorithms and SSD (Single Shot Multibox Detector). They are commonly used for  real-time object detection as, in general, they trade a bit of accuracy for large improvements in speed.

Two Main Groups of Object Detection Algorithms

Dive into Deep Learning

This algorithm uses a single convolution neural network but instead of processing the image pixel by pixel and developing more broad features with every step in the hidden layer. It takes in the whole image and performs the object detection in a single regression problem. It first splits up the image into an SxS grid before passing it into the neural network where bounding boxes and class predictions are then created. Intersect over Union and non-maximum suppression are needed to formulate the founded boxes.

YOLO (You Only Look Once)

The single-shot multibox detection (SSD) model is a multiscale object detection framework that operates in a single forward pass. It primarily consists of a base network followed by a sequence of multiscale feature map blocks. The architecture generates varying numbers of anchor boxes at different sizes across these blocks, and detects varying-size objects by simultaneously predicting the classes and spatial offsets of these anchor boxes.

Learn Before

Related

Learn After