A pre-trained model is the one that has been trained on a previous problem and that can be used to solve other problems of similar domains. The architecture of these models can be slightly altered so that the model can be fine-tuned as per the requirements of the application it is being used for. Pre-trained or Transfer Learning models have gained huge popularity for its implementation in Object Detection based applications. Here are some of the very popularly used pre-trained object detection models listed in no specific order.
1. R-CNN
R-CNN uses search selective method to find the regions to detect objects after it passes through convolutional networks. It uses 2k areas that are passed to aa SVM classifier and then the ground truth-truth boxes are computed. It is highly efficient as it passes the image to the pre-trained CNN only once. The input images are used to predict the detected object. Bounding box localisations are used along with linear regressor to propose the search selective process in R-CNNs.
2. Resnet50
The Resnet50 is a deep residual neural network that can also be used for object detection. It is faster and efficient. Resnet50 along with ImageAI can be used to develop object detection applications. Resnet50 is deeper and provides better training and testing capabilities. It also has other variants such as Resnet101 etc. Resnet is also used in many other applications like Image Classification, image masking, etc.
3. FPN
It uses pooling as a method of selection. It contains a single feature map and a pyramid based feature hierarchy, is what makes it useful by carrying out end to end training. They also tend to obtain more accurate results than other methods. It is highly focused on building feature pyramids in the ConvNets. FPN uses pre-trained convnets for multi-scale testing to create feature pyramids. The use of FPN has shown a great improvement in the feature extraction in many recognition systems.
4. Retinanet
Retinanet is a state of the art Transfer Learning based Neural Network approach used for Object Detection based applications in deep learning. Retinanet improves the prediction accuracy and resolves the class imbalance problem. Thus, inhibiting around 100,000 boxes. Retinanet has a upper hand in achieving higher accuracy in detecting objects. Retinanet was initially created to overcome the shortcomings of the SSD and YOLO based neural networks.
5. Yolo V3/V2
Yolo or ‘You Only Look Once’ can detect over 9000 object categories. A basic YOLO model has 2 categories. Yolo V3 is the latest version of in the Yolo object detection series. Yolo is simple to understand as it takes an input image and then learns the class probabilities instantly. It only runs the input image once through the CNN. Yolo is one of the most used Transfer Learning approach in Object Detection applications. Yolo also works on unlabelled data also. So, Yolo is super fast and highly accurate than its peers. Yolo uses custom functions that cater to increasing the stability of the overall model.
6. Faster R-CNN
It is similar to the R-CNN approach but the only difference here is the induction of convolutional feature map that is used to detect the regions in an image to reshape them into a bounding box. It profoundly replaces the Selective Search technique with much efficient Region Proposal Network that generates the detected areas in an image. It is highly used for carrying out real-time performance-based tasks in object detection tasks. It provides a trade-off between accuracy and speed. Faster R-CNN was the first to include the anchor boxes. It also carries out processes like testing and training way faster than R-CNN.
7. SSD
SSD or Single Shot Detector is a multi-box approach used for real-life object detection. This technique is built on VGG-16 architecture. It takes only a single shot to detect multiple images within a given input image. It is different than other models on this page as a pre-trained model like RNN takes more than 1 shot to detect objects in an image. SSD is easy to train and integrate the detection components into the object detection system. SSD is one of the most popular approaches used today.
Final Words
These Transfer Learning Models are efficient alternatives to the lengthy process of creation of a deep learning Object Detection Model From Scratch. Transfer learning approaches are primarily used for solving a new problem by using old problem results. Any new dataset can be used for training a pre-trained model after concurrently fine-tuning the pre-trained model. Pre-trained models are also very useful where the data available is not quite large. Generally, it needs to be smaller than the one used to train the original pre-trained model. By using pre-trained models we are able to retain the functionalities, features, and weights of the previous model when utilizing it another one.
Share Your Views: