SYSTEM AND METHOD FOR MODELING A THREE DIMENSIONAL SPACE BASED ON A TWO DIMENSIONAL IMAGE

摘要

A system and method for automatically modeling a three-dimensional space based on a two-dimensional image of the three-dimensional space, by obtaining line segments corresponding to boundaries between surfaces of a three-dimensional space which are visible in a two dimensional image, identifying super-pixels and super-pixel boundaries between the super-pixels in the two-dimensional image of the three-dimensional space, and assigning a first weighting value to each identified super-pixel boundary. The first weighting value assigned to each super-pixel boundary is determined by assigning a first likelihood value to each super-pixel boundary based upon the perpendicular distance between the super-pixel boundary and the nearest line segment, assigning a second likelihood value to each super-pixel boundary based upon the difference in orientation between the super-pixel boundary and the nearest line segment, calculating the product of the first likelihood value and the second likelihood value, and assigning the product as a first weighting value to each pixel of the super-pixel boundary. Estimating the probability of each pixel of the two-dimensional image representing a boundary between two objects in the two-dimensional image and assigning a second weighting value to each pixel based on the estimated probability, determining a third weighting value for each pixel of the two-dimensional image wherein the third weighting value is the product of the first weighting value and the second weighting value. Partitioning the graph formed by the determined third weighting values with a number of partitions corresponding to the number of surfaces of the three-dimensional space which are visible in the two dimensional image to obtain a partitioned graph, combining the partitioned graph with the second weighting value to obtain a super-pixel graph by summing the third weighting value and the second weighting value assigned to each pixel, and segmenting the super-pixel graph.