I've been playing with transfer learning lately. Originally I expect to have to label thousands of images manually to train object detection on new classes, but I found it easier to bootstrap with synthetic data (just drawing a small number of images from a very limited set), find the worst predictions and label them manually, repeat a few times, while retraining. I only had to complete this cycle on a few tens of images before I had an awesome object detector that generalized to a wide range of image conditions.