On the Performance Evaluation of 3D Reconstruction Techniques from a Sequence of Images
- PDF / 1,420,842 Bytes
- 8 Pages / 600 x 792 pts Page_size
- 44 Downloads / 173 Views
On the Performance Evaluation of 3D Reconstruction Techniques from a Sequence of Images Ahmed Eid Computer Vision and Image Processing (CVIP) Laboratory, University of Louisville, Louisville, KY 40292, USA Email: [email protected]
Aly Farag Computer Vision and Image Processing (CVIP) Laboratory, University of Louisville, Louisville, KY 40292, USA Email: [email protected] Received 1 January 2004; Revised 20 September 2004 The performance evaluation of 3D reconstruction techniques is not a simple problem to solve. This is not only due to the increased dimensionality of the problem but also due to the lack of standardized and widely accepted testing methodologies. This paper presents a unified framework for the performance evaluation of different 3D reconstruction techniques. This framework includes a general problem formalization, different measuring criteria, and a classification method as a first step in standardizing the evaluation process. Performance characterization of two standard 3D reconstruction techniques, stereo and space carving, is also presented. The evaluation is performed on the same data set using an image reprojection testing methodology to reduce the dimensionality of the evaluation domain. Also, different measuring strategies are presented and applied to the stereo and space carving techniques. These measuring strategies have shown consistent results in quantifying the performance of these techniques. Additional experiments are performed on the space carving technique to study the effect of the number of input images and the camera pose on its performance. Keywords and phrases: evaluation, 3D reconstruction, reprojection, quality assessment, stereo, space carving.
1. INTRODUCTION The design of experimental test beds and methods of analysis in addition to the definition of the ground truth are important components of any performance evaluation system in computer vision. Furthermore, the performance must be quantified to make valuable objective comparisons. Otherwise, the failure of assessing the performance may not only cause unnecessary complexities in the subsequent processes but also may lead to the inability to fulfill the requirements of the application under concern. This reflects the need of a general methodology for testing the performance of similar computer vision techniques. Particularly, in this paper, we address the problem of performance evaluation of 3D reconstruction techniques from a sequence of images. In fact, this area of research still faces a lack of standardized and widely accepted methods. Yet, the seminal works of Szeliski and Zabih [1, 2] are leading examples in the sense of presenting new metrics and methodologies of performance evaluation of stereo and motion techniques. Another related work [3] has presented a design of an experimental setup for performance evaluation of stereo techniques in telepresence. This setup uses a 3D scanner for providing the ground truth for performance evaluation of stereo techniques.
In this paper, we extend the evaluation domain to include
Data Loading...
