I have a model which detects an object and makes a bounding box over it. The problem is that those bounding boxes are not accurate and need to be a little more tight on the object rather than some body parts exceeding the box or some boxes bigger than the object size. I want to apply reinforcement learning to make bounding boxes more accurate as I have the information of perfect bounding boxes which is the target and the input images which have the inaccurate bounding boxes or the inaccurate coordinates. I found a paper online on the exact same topic but I cant find the code for the it builds an environment with defined states, actions and awards. As I am very new to reinforcement learning I can not make the environment from scratch.
Here is the paper https://melaniemitchell.me/ResearchGroupContent/MastersTheses/AndrewClelandThesis.pdf
Is this whole approach of using and changing grid size measures as states and actions doable? If yes then can someone please link me to a code preferably on github which builds quite similar environment? If not then can someone give any suggestion of building the environment or what other approach I could use?
There is a similar paper:Multi-stage Reinforcement Learning for Object Detection which does the same thing.
The code for the implementation of the paper can be found here.