Performing structure from motion for unordered images of a scene with multiple object instances

A system for performing structure from motion from scenes with large duplicate structures, comprising:

• a mass storage device to store images obtained from a camera taken during a single time period;
• a match graph computation module to obtain a pairwise match graph generated by using interest point detection in the images to enable pairwise matching of a set of the images;
• a scene reconstruction module to estimate multiple metric two-view and three-view partial reconstructions by performing independent structure from motion computation on a plurality of image match-pairs and image match-triples selected from the pairwise match graph;
• a classification module to classify a set of image match-pairs selected from the pairwise match graph into a first subset of correct matches and a second subset of erroneous matches using an expectation maximization operation and a scoring function to evaluate a plurality of match labeling hypotheses;
• a global geometry module to apply a global geometric consistency analysis to the image match-pairs;
• and
• a structure from motion module to perform a structure from motion computation on the first subset of correct matches which have been inferred as correct matches as a result of the classifying.

A method for performing structure from motion for unordered images of a scene with multiple object instances, comprising:

• obtaining a pairwise match graph using interest point detection in the images of the scene to enable pairwise matching of the images;
• estimating multiple metric two-view and three-view partial reconstructions by performing independent structure from motion computation on match-pairs and match-triplets selected from the pairwise match graph;
• classifying a set of pairwise image matches identified based on the pairwise match graph into a first subset of correct matches and a second subset of erroneous matches by:
• (a) deriving a plurality of match labeling hypotheses for the set of pairwise matches in the pairwise match graph using an expectation maximization operation which jointly refines a plurality of camera pose estimates based on pairwise pose measurements;
• and
• (b) using a scoring function to evaluate the match labeling hypotheses to identify one of the match labeling hypotheses with a highest score that is a final solution;
• and
• performing a structure from motion computation on the first subset of correct matches which have been inferred as correct as a result of the classifying.