《基于贝叶斯估计的2D特征评价》青少年教育文学丛书.pdfVIP

Performance Evaluation of 2D Feature Tracking Based on Bayesian Estimation Yan Li Liu Wenyin Heung-Yeung Shum Microsoft Research China, 49 Zhichun Road, Beijing 100080, PR China {yli, wyliu, hshum}@ Abstract. Feature tracking methods based on Bayesian estimation are widely studied in computer vision systems. The performance of Bayesian decision, however, remains an open problem because an implementation of Bayesian estimation is significantly affected by many parameters in modeling the prior and observation probabilities. In this paper, we evaluate the performance of our MAP based feature tracking algorithm with various parameter settings for many features. For most 2D feature points in our experiments, we found that the uni- form distribution model (or Gaussian model with a very large variance) with linear prediction yields the best feature tracking performance. 1 INTRODUCTION Bayesian estimation or its most common form—maximum a posteriori probability (MAP) estimation—has been widely used in many applications of computer vision and pattern recognition [1][4]. However, its performance is seldom evaluated objec- tively and quantitatively, despite that many applications are claimed successful. Usu- ally, the probability densities involved in the MAP-based solutions have to be mod- eled differently for different applications. At the very least, several parameters should be decided/set differently for different applications on different test (input) data. Al- though the Bayesian estimation method (and its basis—the Bayes Theorem and Bayes decision rule) is rational in theory, it is generally hard to find a common set of parameter values that make the method work at the same l