Teste de software para sistemas de reconstrução 3d baseada em imagens

Abstract

This work presents an automated testing approach for 3D reconstruction systems based on images, with a focus on the quality and reliability of these systems. 3D reconstruction systems are utilized in various fields, including medicine, engineering, and entertainment. However, due to the complexity of the inputs (sets of images) and outputs (3D models), the testing process becomes challenging, especially regarding the role of test oracles, which are responsible for validating the system’s correctness. In this context, this work develops an automated test oracle using deep neural networks, aiming to validate the results of the generated 3D models. A literature review identifies several challenges in the area, particularly the lack of testing in the 3D reconstruction domain. Related domains, such as image processing, computer vision, and computer graphics, are also considered. The approach encompasses everything from the selection and preprocessing of input and output images to the detailed evaluation of the generated 3D models, using deep learning techniques to compare patterns between the models and input images. The COLMAP software serves as the basis for validating the approach, and defects are intentionally introduced to analyze the oracle’s behavior when handling failures. The automation of tests aims to ensure standardization, time savings, repeatability, and reliability in the process. As a result, the approach contributes to advancements in the software testing field, supporting the development of higher-quality 3D reconstruction systems.