Examination Committee

Prof Michael K Y WONG, PHYS/HKUST (Chairperson)
Prof Bertram E SHI, ECE/HKUST (Thesis Supervisor)
Prof Ling SHI, ECE/HKUST

Abstract

Joint development of perception and behavior based on the active efficient coding framework has been shown to be an effective approach to study the development of eye movements. Disparity vergence, one type of eye movement, is believed to involve cooperation between fovea and periphery. In realistic environments, objects at different depths project conflicting disparity information to different regions of retina.  This cannot be solved effectively by the previous developmental models of vergence control, where information from fovea and periphery as pooled together with same weights.

This thesis proposes a hierarchical approach to integrating foveal and peripheral vision for vergence control. The model consists of two levels.  Three lower level policies receive input from the fovea, inner periphery, outer periphery of the retina, respectively, and generate separate vergence commands.  One top-level policy learns an effective mechanism of choosing one out of the three options for vergence control in different situations.

The hierarchical model for integration of fovea and periphery is proved quantitively to have more accurate and robust vergence control over previous model in the realistic testing environments. In addition, it shows good performance in dealing with some extreme situations, such as tracking small object moving in depth, lack of context in fovea, monocular input in fovea due to occlusion. Furthermore, a periphery to fovea process is observed when applying the hierarchical model in vergence control, which is consistent with psychological evidence found by other researchers.

Keywords- stereo disparity; vergence mechanism; reinforcement learning; multiple agents; fovea and periphery