With respect to object detection training of two object classes with one class visually less distinct than the other: improved the less distinct class’s detectability by initially training on annotations of that class only and then after achieving a stable map, introducing the annotations of the other class in the training.
