The speed and accuray of tiny model

The tiny model such as yolov3-tiny or yolov4-tiny is designed for limited computing resource.Although their speed increases much, the accuracy decrease obviously. If there is a good way to improve speed and accuracy jointly?

This sounds like the goal of a lot of research projects as it’s a win-win. :wink:
Besides general good research ideas and clever model architectures I don’t know if there are other recommendations.