Aerial imagery has grown to sizes and resolutions beyond comprehension. For Detroit alone, Maxar satellite imagery is around 50 GB with aerial imagery being around 100 GB.

Inference on this scale of imagery is expensive, but paired with an Objectness Activation Network (OAN), a much cheaper threshold for each tile is calculated to direct the pipeline on whether a tile has information at all.

Using this technique I have increased the speed of inference by almost a third without loss in accuracy in areas where features are not as prevalent.