in this project we aim to address vehicle detection (classification and localization) challenges by fine tuning three Deep Learning models : R-CNN , YOLOv5, RetiaNet
The initial approach used in this project is the R-CNN model, which is a relatively simple model compared to others. However, it has limitations in terms of accuracy and implementation time. The selective search algorithm, responsible for object candidate box allocation, consumes more than 90% of the execution time. This issue is particularly problematic for real-time detection applications, making the use of this network less favorable.
To address these challenges, the problem is reframed as a classification problem. The project selects parts of the image that include the object of interest (e.g., a car) and parts that do not. Additionally, the model includes a background category. Figure bellow shows an example from the Vedai dataset where an image is classified as a vehicle during the R-CNN model's training process.
After obtaining training data, a common convolutional network like VGG is used for classification. During testing, each image goes through the selective search algorithm to generate proposed boxes. These boxes are then fed into the model for prediction. If the model identifies a specific box as non-background, it is reported as containing an object.
Overall, this approach aims to improve accuracy and reduce implementation time by redefining the problem as a classification task and optimizing object candidate box allocation using selective search algorithm.
In the next phase of the project, we explored the RetinaNet model, which is a fascinating and powerful network known for its unique hierarchical structure. This network strikes a great balance between accuracy and speed, performing on par with the Faster R-CNN in terms of accuracy while maintaining a speed comparable to real-time networks. For this project, we utilized the Retina model with transfer learning, and the evaluations pertaining to this model will be discussed in the following section of this chapter. To provide visual representation, a portion of the output from the RetinaNet component is displayed in the figure below. This figure showcases the output of the Retina model alongside its corresponding prediction labels for some of the test data.
The image above illustrates the results of the Retina model on the test data. In this image, the green box represents true positives (TP), the yellow box represents false positives (FP), and the red box represents false negatives (FN).
The Yolo architecture differs from previous networks that divided images into different regions to check the probability of object presence in certain regions. Yolo performs training on the entire input image using a convolutional neural network and can predict multiple boxes for each image. The input image is divided into S × S parts, and if an object's center falls on one of these parts, that part is responsible for detecting the corresponding object. Each part predicts the number B of the box, which includes the probability of object presence in that part of the box, as well as the defining coordinates of the box. This architecture defines each square by six numbers: pc, bx, by, bh, bw, and c. Pc represents the confirmation coefficient of object presence in that part, while c is a vector indicating the probability of that object belonging to different classes. Pc and c are related in that pc indicates the probability of object presence in a certain part, and c indicates the probability of the object belonging to different classes. The Yolov5 model is a member of the Yolo family and is suitable for real-time applications due to its unique architecture. It offers several options, ranging from small models with few parameters to models with many parameters suitable for working on complex and extensive data.
An example of Yolo model output on test data: The model has performed much better than the Retina network in recognizing more difficult data.
In this section, we conducted an evaluation of various models used in this project. Choosing the best model for a given dataset and real-world applications can be a challenging task, and it requires careful consideration of factors such as accuracy and speed. In some cases, even minor changes in the model's parameters can lead to a significant reduction in accuracy. Conversely, larger changes that improve speed may not have a significant impact on accuracy. Therefore, it is essential to strike a balance between these two factors when selecting a model.
Another challenge is finding the right parameters for batch size and the number of training rounds. These parameters can have a significant impact on the model's performance and should be carefully considered.
here, we explored various techniques, such as transfer learning and testing models with different sizes, to make them more suitable for real-time applications. We also investigated other factors that could affect the model's accuracy and speed. The results of our evaluation and comparison of different models are presented in Table below. It is worth noting that the region-based model, R-CNN, was not suitable for object detection due to its poor performance on the available training data. Therefore, we compared Retina and Yolo models instead.
