Acquiring Accurate Visual Information from Images Using Border Ownership
Project No: 111E155
Project Duration: 3 years (01.04.2012-01.04.2015)
As our first goal, we will collect human-labeled images that contains the border ownership of the edges. Using this labeled data, we will investigate the relationship between different visual information and the border ownership assignment. We argue that such an investigation is crucial for a problem whose underlying mechanisms are unkown (like border ownership) since it is known that human vision system utilizes the regularities in the scenes and investigating such regularities (from labeled data) can yield important facts about different visual processes.
As our secold goal, we will use using tensor voting method to model the visual modalities (that is involved in border ownership) and the interaction between these modalities. We find tensor voting as a suitable method for our computational modeling purposes since it has been demonstrated to be useful for modeling different information sources and their interactions, which, we expect, will be the case based on our investigation in the first goal.
As our final goal, we will apply the computational model derived in our second goal to three important vision problems: optic flow and stereo disparity computation, and depth estimation at image regions from the edges in the scene. As known, using local mechanisms for optic flow and stereo disparity computation leads to incomplete and ambiguous visual information; therefore, these three applications will be useful for testing our computational model. Moreover, we will test whether using border ownership with stereo helps a humanoid robot in grasping objects with textures of different complexity.