It is also possible to train a model to detect objects that are not in the COCO dataset. To do so, a large amount of data is needed. In general, it is recommended to have at least 1,000 samples per object class. To generate a training set, training images need to be manually annotated by drawing the bounding boxes around them.
Using the Object Detection API does not involve writing Python code. Instead, the architecture is defined using configuration files. We recommend starting from an existing configuration and working from there to obtain good performance. A walk-through is available in this chapter's repository.
