For the foveated perception module, I think it would make sense to have a discrete selection of where the agent can "look" next. This would make it easier to learn, although the bounds will have to change depending on the size of frame. Currently, a change in both the x and y direction are sampled. This should be changed to having the agent select two coordinates to serve as the top left corner of a fixed size square.
I'll have to do some reading about how this is done for bounding box methods.
For the foveated perception module, I think it would make sense to have a discrete selection of where the agent can "look" next. This would make it easier to learn, although the bounds will have to change depending on the size of frame. Currently, a change in both the x and y direction are sampled. This should be changed to having the agent select two coordinates to serve as the top left corner of a fixed size square.
I'll have to do some reading about how this is done for bounding box methods.