Kuijper, ArjanKnauthe, VolkerChahota, Karanveer SinghKaranveer SinghChahota2023-11-022023-11-022023https://publica.fraunhofer.de/handle/publica/452587In computer vision, object detection plays a very important role in the field of image processing. Therefore, lots research in this topic, especially for opaque objects, are available utilizing different methods, such as machine learning, to increase the accuracy and robustness of those detectors. The detection of transparent objects, however, still poses a unique challenge in recent research. In order to detect the unique material properties, such as reflection or refraction, often machine learning methods are used. However, it is difficult to interpret the decision making processes of a machine learning methods due to the complexity and the black-box nature of those methods, such as in large deep neural networks. There are some scarce research about transparency detection approaches using only traditional computer vision methods and physical cues of the transparent objects, with the goal of creating a transparency detection method without using machine learning methods. Using only physical cues of one image, however, makes it difficult to deliver perfect results. With the usage of multi-view images, those results can be refined, utilizing the global knowledge of the scene with different perspectives. In this work, a pipeline is proposed that uses a transparency detection algorithm, which is based on region similarity metrics of single-view RGB images, and utilizes the advantages of multi-view images to refine the results. First, synthetic datasets were generated in order to analyze the applicability of the transparency detection on synthetic data. Then the utilization of common 3D reconstruction methods are analyzed and discussed. Finally, we propose an approach utilizing the transparency detection algorithm with multi-view data and geometric hashing and discuss the benefits and opportunities of this approach.enBranche: Information TechnologyBranche: Cultural und Creative EconomyResearch Line: Computer graphics (CG)Research Line: Computer vision (CV)LTA: Machine intelligence, algorithms, and data structures (incl. semantics)LTA: Generation, capture, processing, and output of images and 3D models3D Data acquisitionTransparency computationImage segmentationmaster thesis