Target detection model regression anchor offset and other problems

Why learn offsets instead of actual values ？

  Anchor Has been roughly “ It's framed ” Enter the target in the image , The obvious problem is that ： The frame is not accurate enough . Because limited by Anchor How to generate ,Anchor The coordinates of are always fixed . therefore , If we need to predict relative Anchor Of offset, that , Can be predicted by offset Adjust the anchor frame position , So as to get more accurate bounding box.

Why learn the offset coefficient instead of the offset ？

   because w and h All positive numbers. , When the last layer of the network returns to the output, it cannot guarantee the positive and negative , So the simplest way is to calculate the predicted output exp, This ensures that the predicted value is always positive , So in turn , The goal of prediction is to seek log