| An objective evaluation framework for range image segmentation has been created and utilized in developing an improved range image segmentation algorithm. The framework automatically trains an algorithm by using training sets and compares the performance of two algorithms on test sets by using statistical test. We adopted the University of Bern curved-surface segmentation algorithm as the algorithm to be analyzed for improvement. This algorithm extracts edges and regards an area enclosed by a connected edge contour as a potential region. Since a typical edge extraction has gaps along true edges, most initial regions need to be split into smaller ones. For each initial region, the baseline algorithm hypothesizes a quadratic surface. If the surface fit error is below predefined thresholds, the region is accepted. Otherwise, the region should be recursively split until every region satisfies its own surface hypothesis. The split is performed by dilating every edge pixel inside the region to link edge segments. But this may produce false edge contours by linking noisy edge points. The new approach presented in this dissertation inspects how those unsuccessful regions fit the hypothesized surfaces and then performs operations according to the fitting patterns, rather than linking edges blindly. The experimental results show that the new approach produces better performance. The whole procedure of evaluation of the new algorithm, including training, validation, test, and comparison to the baseline algorithm was performed using the evaluation framework. The experimental work done in this dissertation shows that the evaluation framework is useful in developing new algorithms or verifying improvement of an existing algorithm. |
