CVAT Image Annotation Tool

In the last few chapters, we exercised our skills with object detection and semantic segmentation CNNs by downloading pretrained weights and using them as is. We’ll probably never do this in a real-life automated inspection project unless working on an application that detects cats, dogs, and pedestrians in images of urban scenes.

A more realistic scenario would imply transfer learning or fine-tuning a pretrained CNN on an image dataset of the objects that matter for your inspection task. That means building our dataset with a sample of images we want to inspect, including examples of the defects, if applicable.

In the case of a classification task, building our annotated dataset can be as simple as grouping images of objects of the various categories in separate directories. In the case of object detection or semantic segmentation, creating the dataset is more complicated. The object’s bounding box for object detection or the mask for semantic segmentation is typically drawn by hand using a graphical user interface. We must save this data in a file.

Many tools are available to annotate images. The one we’re going to explore is CVAT, a free, open-source image ...