Metrics for object detection problems
In recent years, deep learning competitions have become more and more common on Kaggle. Most of these competitions, focused on image recognition or on natural language processing tasks, have not required the use of evaluation metrics much different from the ones we have explored up to now. However, a couple of specific problems have required some special metric to be evaluated correctly: those relating to object detection and segmentation.
Figure 5.4: Computer vision tasks. (Source: https://cocodataset.org/#explore?id=38282, https://cocodataset.org/#explore?id=68717)
In object detection, you don’t have to classify an image, but instead find relevant portions of a picture and label them accordingly. For instance, in Figure 5.4, an object detection classifier has been entrusted to locate within a photo the portions of the picture where either dogs or cats are present and classify each of them with a proper label. The example...
