Learning from Rich Semantics and Coarse Locations for Long-tailed Object Detection