I have to locate cars in satellite images. Those images are some 10,000x20,000 pixels and the car only take about 15x15 pixels, being mere shadow with minimum texture. I managed to extract the roads, where cars are mostly found, so I limited my search area to the roads. In short I have a two-class problem, being "car" and "street" (or "car" and "no-car"). Deep learning is good for separating even hundreds of classes, but with more texture I guess. What would be the best strategy for my problem?
Thank you in advance,