How much do I have to know about a problem to solve it using computer vision?

In fact, surprise it may seem, not much. All you have to know is the basics on how to train those models. For instance, you can train a model to separate viral from bacterial pneumonia, without actually knowing the difference. It took me a while to learn minimally the difference between false and true coral snakes, even though the model learnt much faster than me. Of course, if you have knowledge on the area of the image, it is beneficial, but not requirement.

The nice feature about those models is precisely that. If you have the images well-annotated by experts, and you can find them generally online, you can train the model, and the model will be smarter than you in no time. You can share the model.

Machine intelligence, different from humans, can be easily shared after the model was trained. You can for instance share you model as JSON on GitHub. The model is a set of numbers, the network weights. All the images with annotations became weights, which can be reused by anyone.