The model:

The model uses operators that model the responses of cells in MT. They are tuned to the direction of motion within their receptive fields and have an excitatory and inhibitory region. These operators essentially subtract the motion from adjacent regions of the visual field.

Image of operators:

Each region of the visual field is processed by a group of these operators that differ in their preferred direction of motion and the angle of the axis between the excitatory and inhibitory regions.

Image of model:

The maximally responding operator at each location projects to the next layer of cells. These cells are tuned to a radial pattern of input, in terms of the preferred directions of motion, from the motion subtraction operators. The maximally responding cell in this layer will have a center of the radial pattern that coincides with the translation direction of motion of the observer.

Image of model:

The model works by removing rotational components of image velocity, leaving only the image velocity due to observer translation. The preferred directions of the maximally responding motion-subtraction operator across the visual field form a radial pattern.

Image of flow field for translation and rotation through 3D cloud:

Image of preferred directions of operator responses:

The model computes heading well in the presence of rotations.

In addition, the magnitude of the motion-subtraction operator's response is proportional to the change in depth between the excitatory and inhibitory regions.  Thus, we can detect the location and magnitude of depth changes in the scene:

Image of operator response magnitudes for wall in front of plane. 

Added observer rotations do not affect this response:

Image of operator responses in the presence of rotations:

The responses at a depth edge are proportional to the difference in inverse depth between the two surfaces.

Image of graph of responses: