It is important to keep bounding boxes as close to the object you want to identify as possible. This method will help you avoid capturing unimportant areas, like the background. During model training, the model will look at every pixel inside a bounding box. If you include extra, unimportant pixels, the model might think these pixels are also important. This may cause complications when the model is deployed and is looking for objects in the real world.For example, in the image below, the two wind turbines are labeled with one bounding box. While this approach successfully captured both wind turbines, it also included a significant amount of background.The image below shows an improved labeling strategy where two bounding boxes are used to label each wind turbine. This method ensures a more focused and accurate representation of wind turbines, minimizing the inclusion of irrelevant areas.
You can adjust the size of a bounding box to more accurately identify the object of interest. Remember you want to draw bounding boxes as close to the objects of interest as possible.To resize a bounding box:
Open the image that has the label you want to resize.
Move your cursor to the corner of the bounding box that you want to adjust. Your cursor will display a double-sided arrow cursor (resize cursor).
Click and drag the bounding box until you see the results you want.