On Creating Reference Data for Performance Analysis in Image Processing

Abstract

This thesis investigates methods for the creation of reference datasets for image processing, especially for the dense correspondence problem. Three types of reference data can be identified: Real datasets with dense ground truth, real datasets with sparse or missing ground truth and synthetic datasets. For the creation of real datasets with ground truth a existing method based on depth map fusion was evaluated. The described method is especially suited for creating large amounts of reference data with known accuracy. The creation of reference datasets with missing ground truth was examined on the example of multiple datasets for the automotive industry. The data was used succesfully for verification and evaluation by multiple image processing projects. Finally, it was investigated how methods from computer graphics can be used for creating synthetic reference datasets. Especially the creation of photorealistic image sequences using global illumination has been examined for the task of evaluating algorithms. The results show that while such sequences can be used for evaluation, their creation is hindered by practicallity problems. As an application example, a new simulation method for Time-of-Flight depth cameras which can simulate all relevant error sources of these systems was developed.