A pointwise smooth surface stereo reconstruction algorithm without correspondences☆
Graphical abstract
Highlights
► 3D surface reconstruction from 2D point sets without correspondences. ► Surface extraction from space of all potential epipolar reconstructions. ► Application to breast cancer screening.
Introduction
Why write another article on yet another method for 3D reconstruction from digital image pairs? This article addresses a specific type of problem that may be encountered in some applications for which standard techniques are not appropriate or are difficult to use. The problem is that of reconstructing in 3D a relatively dense set of specific points on a smooth surface from two calibrated views, but in a situation where solving the correspondence problem is difficult due to a high degree of self-similarity between the image features.
The motivation for this research came from a problem in medical imaging. In the DIET (Digital Image-based Elasto-Tomography) project [1], it is necessary to reconstruct in 3D the motion of a large set of points on a human breast surface as it is being mechanically vibrated. The motion of specific points on the surface is required, rather than a description of the surface motion as a whole, because this point motion is required to solve an inverse problem for the 3D distribution of internal elasticity of the breast tissue.
Because in this type of application, specific feature points are required that can be tracked as the surface deforms, volume-based methods [2], [3], [4] based on visual hulls that do not measure the location of trackable individual surface points are not appropriate. Likewise, state of the art stereo methods such as graph cuts [5], [6] and belief propagation [5], [7] that are based on Markov Random Fields are also inappropriate, as they are concerned with creating a 3D reconstruction of a scene by computing a pixel-by-pixel depth map, and once again do not compute the precise locations of trackable individual surface points. For the DIET problem small artificial fiducials are randomly applied to the breast surface. For practical reasons these features are essentially identical in appearance, and hence difficult to correspond using standard interest-point detectors [8], [9], [10]. Dense stereo methods [11], [12] relying on accurate correspondence of a number of key features are therefore also not well-suited to this application. Dellaert et. al. [13] have developed an algorithm for computing feature point correspondences without any a priori information by computing the maximum likelihood estimate of the scene and cameras using an EM algorithm. This method is potentially a promising alternative approach, however significant work would be required to modify the algorithm to incorporate large numbers of occluded features and an arbitrary number of world features.
The types of application envisioned for the algorithm presented herein are those where a (relatively) dense set of self-similar points on a smooth surface needs to be reconstructed in 3D from fully calibrated cameras. Another application might be in surveying, for example, where landmarks such as trees or plants are the feature points, and the goal is to reconstruct the topography on which they lie from multiple aerial views.
Section snippets
Camera model
The computer vision notation used follows that of Hartley and Zisserman [14], a standard reference for multiple view geometry. In this paper, cameras are fully calibrated, and a 3 × 4 camera projection matrix P is assumed to be of the formwhere is a rotation matrix, a 3-vector, and an upper triangular matrix representing the internal parameters of the camera. The projection, in homogeneous coordinates, from world coordinates X = (X, Y, Z, W)┬ to image coordinates u = (u, v, w)┬ is
Problem Definition
Let be the surface in Euclidean space to be reconstructed and let be the set of feature points on . The feature points are points on the surface that can be reliably extracted and localised in images, but are indistinguishable in appearance from one another. Make the following two assumptions about : Assumption 1 Smoothness The surface has continuous directional derivatives, and the maximum principal curvature over the entirety of is bounded by some known constant κ Assumption 2 Feature distribution The feature points are
Test surface
A number of experiments were performed using simulated data. For a test surface, the functionwas used. Points were randomly generated on the surface by choosing x and y from U(− 0.5, 0.5) and generating the corresponding z values. This gives a slight irregularity to the surface, but this effect is not large. This function has maximum principal curvature given by
Implentation Details
All code in this section was written purely in MATLAB and experiments were run on a single 1.2
Discussion
The results in the case study were produced with a MATLAB implementation of the algorithm, executed on a 2.53 GHz dual-core laptop with 4 GB RAM. The MATLAB version was R2011a in a 64-bit Linux environment. With this implementation, the full surface reconstruction procedure for the DIET system takes around 11 seconds, not including the extraction of the point features by threshold segmentation. There are numerous optimisations, not implemented here, that could be made to improve the efficiency of
Acknowledgements
Thanks to the Tertiary Education Commission (TEC) for providing funding through a Top Achiever Doctoral Scholarship for RGB. Thanks also to the anonymous referees for constructive reviews and for suggesting references and improvements to the content and presentation.
References (18)
- et al.
Digital image-based elasto-tomography: first experiments in surface based mechanical property estimation of gelatine phantoms
JSME Int. J.
(2005) The visual hull concept for silhouette-based image understanding
IEEE Trans. Pattern Anal. Mach. Intell.
(1994)- et al.
Polyhedral visual hulls for real-time rendering
- et al.
Hardware-accelerated visual hull reconstruction and rendering
- et al.
Comparison of graph cuts with belief propagation for stereo, using identical MRF parameters
- et al.
Multi-camera scene reconstruction via graph cuts
- et al.
Stereo matching using belief propagation
Distinctive image features from scale-invariant keypoints
Int. J. Comput. Vision
(2004)- et al.
SUSANA new approach to low level image processing
Int. J. Comput. Vision
(1997)
Cited by (11)
Silicone phantom validation of breast cancer tumor detection using nominal stiffness identification in digital imaging elasto-tomography (DIET)
2018, Biomedical Signal Processing and ControlCitation Excerpt :The cameras and actuator are synchronized by strobing a set of LEDs at specific phases in the cycle. Consecutive sets of 2D images are converted into a 3D description to measure the surface displacement of each reference point between frames [21,28]. The DIET system is intended to be compact and can be easily deployed, event in remote areas, where people have little access to the breast cancer screen facilities [27].
A Surface Vibration-based Method for Tumor Detection of Women Breast in a DIET System
2017, Procedia EngineeringStructural health monitoring of tissue mechanics for non-invasive diagnosis of breast cancer
2018, At-AutomatisierungstechnikFinite element modelling and validation for breast cancer detection using digital image elasto-tomography
2018, Medical and Biological Engineering and Computing
- ☆
This paper has been recommended for acceptance by Daniel Rueckert, PhD.