Shape from silhouette probability maps: reconstruction of thin objects in the presence of silhouette extraction and calibration error

Authors: Tabb, Amy

Submitted to: IEEE International Conference on Computer Vision and Pattern Recognition
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/26/2013
Publication Date: 6/15/2013
Citation: Tabb, A. 2013. Shape from silhouette probability maps: reconstruction of thin objects in the presence of silhouette extraction and calibration error. IEEE International Conference on Computer Vision and Pattern Recognition. p. 161-168.

Interpretive Summary: In order to automate dormant pruning of apple, first it is necessary to create a three-dimensional model of trees. In this work, we showed how such a model could be created from the information given by many digital cameras. Since trees reside in a natural environment, errors in the data acquisition process are unavoidable. The method described in the paper was able to create high-resolution three-dimensional models of dormant apple trees even in the presence of various types of error in a laboratory setting.

Technical Abstract: We consider the problem of reconstructing the shape of thin, textureless objects; specifically, leafless trees, when there is noise or deterministic error in the silhouette extraction step or there are small errors in camera calibration. Our approach is voxel-based and casts the reconstruction problem as a psuedo-Boolean minimization problem, where the voxels are the variables of a psuedo-Boolean function and are labeled occupied or empty. Since the psuedo-Boolean minimization problem is NP-Hard for nonsubmodular functions, we provide an algorithm for an approximate solution using local minimum search. Our algorithm treats input binary probability maps (in other words, silhouettes) or continuously-valued probability maps identically, and places no constraints on camera placement. Results are given for three different leafless trees and one metal object where the number of voxels is 54.4 million (voxel sides measure 3.6 mm). Our datasets contain silhouette extraction and camera calibration error, and we use two different types of cameras with different calibration accuracy. Our approach is able to reconstruct complicated branching structure of trees and compares favorably to the visual hull approach.