Object detection is a computer vision technique in which we identify the object and the location of that object in image or video. It is an important part of machine learning and deep learning. In object detection we will learn machine so it will identify the element of an image, like a human. The objective of object detection is to replicate this information using a computer.
It is a main technology for self driving cars or perform pedestrian detection to improve road safety. It also used in many systems like video surveillance or image extraction systems.
Following are the detection algorithms:
R-CNN object detection algorithm
This algorithm proposed by Ross Girshick in 2013, this approach solved the problem of selecting the many numbers of regions.
Architecture
R-CNN use selective search algorithm that generates 2000 region proposals from the image. To compute the CNN features we will provide 2000 images of CNN architecture. Then these features pass to SVM model so we will classify the object that present in region proposal. Finally, we will use a bounding box regression we will use a bounding box regression is to get the bounding boxes for each region.
- Problems
- Speed is very slow because we will select thousands of regions in a single image.
- It will take huge time for training.
- We cannot use this algorithm for for live object detection.
- It requires a very large amount of memory.
- We can not learn model at the time on stage because in this we are using selective search and it is fixed.
Fast R-CNN object detection algorithm
Fast R-CNN proposed by Ross Girshick in 2015 to reduce the problem of R-CNN. In Fast R-CNN we have increased the speed of training and testing the data and also improve the accuracy of object detection.
Architecture
In Fast R-CNN we send the whole image to CNN and get feature map. Then we will take region proposal feature map from the feature map.
With the help of pooling layer we can fix the size of region proposal feature map. Next we flatten this fixed size region proposal feature map. It called feature vector. We use feature vector as input for the final segment. It is a fully connected layer which have two results. First is softmax classification where we determine the class of the object and second is bounding box, in which we get coordinates of object that we found.
Problems with Fast R-CNN
- It also uses selective search algorithms and it is time taking.
- Due to region proposal performance is not accurate.
Faster R-CNN object detection algorithm
- To solve the problem of Fast R-CNN Kaiming introduce Faster R-CNN algorithm in 2015.
Architecture
- In the faster R-CNN we have replaced selective search algorithm with the region proposal network (RPN). RPN makes the algorithm faster. We can also use this algorithm for real time object detection.
- Faster R-CNN used two networks, first used RPN for generating region proposal and network using this proposal to identify the objects.
- The identified region proposals have resized using an ROI pooling layer which is used for classification of the image and forecast the offset values for the bounding boxes.
- RPN time is very less to create a region proposal than selective search.
Advantage
- Speed is very fast.
- More accurate than other algorithms.
- We can use in real time object detection.
Yolo object detection algorithm (You look only once)
Yolo is a powerful algorithm for object detection algorithm. It is proposed by Joseph Redmon in 2015.
Architecture
We will use the full image on a single neural network, it will predict the bounding box and class probability in one evaluation. Because it uses single neural network in whole detection so we can increase the performance of detection. It requires single forward propagation through a neural network to identify the objects.
Advantage
- Yolo algorithm speed is very fast because it will detect objects in real time.
- Accuracy is better than other algorithms.
Disadvantage
- Yolo algorithm not good for detecting small objects.