Author: Zen and the Art of Computer Programming
1 Introduction
Computer vision tasks such as image classification (Classification), object detection (Object Detection), and instance segmentation (Instance Segmentation) all need to distinguish and locate different categories or objects according to the input image. With the deep learning ability of the model and the methods to solve various tasks, people gradually pay attention to the latest technologies in image processing and understanding, such as deep neural network, super pixel, multi-sensor fusion, generative confrontation network, etc. However, how to effectively tune the performance of the CV model to ensure the accuracy, robustness and efficiency of the model has become one of the current research hotspots. In recent years, many interesting new methodologies and optimization strategies have emerged in the fields of deep learning and computer vision, and have achieved good results in performance tuning. This article will elaborate some typical methods and techniques in CV model performance tuning from the following aspects, hoping to help readers better understand these methods and the principles behind them, and use them to improve their own model performance level.
2. Conceptual Terminology
2.1 Target detection frame (Bounding Box)
The target detection frame (Bounding Box) is a rectangular frame used to describe the position and size of the object in the image. The coordinates of the upper left corner are $x_1$, $y_1$, and the coordinates of the lower right corner are $x_2$, $y_2$. As shown below:
As shown in the figure, the target detection frame is mainly determined by four parameters: $x_1$, $y_1$ (coordinates of the upper left corner), $x_2$, $y_2$ (coordinates of the lower right corner).
2.2 IoU (Intersection over Union)
IoU is the ratio of the intersection area of two bounding boxes to the union area. When the two bounding boxes completely overlap, then IoU = 1; when the two bounding boxes do not overlap, then IoU ≈ 0. Computing IoU can be used to measure