Interactive structural illustration — not model outputs, measured features or attention weights.
Current scene
object 1: person object 2: car
Geometry changes shape; image context is still supplied over the entire frame.
Select a tracked object
Inputs for the selected object / historical frame
Selecting another object in the same bucket does not reveal a different object-only spatial map. Objects share the bucket memory; each also has its own pointer vector.
Historical frames selected for spatial attention