visual perception entails solving a wide set of tasks, e.g., object
detection, depth estimation, etc. The predictions made for multiple tasks from
the same image are not independent, and therefore, are expected to be
consistent. We propose a broadly applicable and fully computational m