Research Projects

The aim of this research project is holistic scene understanding in large aerial datasets, consisting of thousands of massively redundant high-resolution images. Holistic scene understanding is one of the major problems in computer vision and photogrammetry and has recently got a lot of attention. The problem of holistic image understanding includes two fundamental tasks: 3D scene reconstruction and semantic interpretation of the imaged content at the level of pixels. The tight interaction between semantic classification and 3D reconstruction is often ignored by state of the art aerial image processing workflows, due to the lack of computational power, the absence of efficient algorithms or the enormous effort of manual intervention. However, these tasks are mutually informative and should be solved jointly as a correct class labelling is a valuable source of information for reconstruction, and 3D information can help to improve the semantic interpretation. For instance, a correct classification is a valuable source of information for reconstruction in regions where dense matching methods fail (e.g. sheets of water and reflecting windows / facades), and 3D information can be used as a prior to improve classification (e.g. building and road detection). The high resolution and redundancy due to large overlaps of aerial images requires massive processing power which will be handled by taking advantage of graphic processing units that have proved to give a significant speedup compared to single core machines. In particular, we will focus on algorithms based on variational methods, which provide a high degree of parallelization capability. In order to reduce cost-intensive manual interaction, we further will exploit publicly available user-data from the Internet to improve both interpretation and 3D reconstruction.

In the HOLISTIC project we will provide a flexible framework for scene classification and 3D reconstruction from aerial images that outperforms current state-of-the art and delivers interpretable models at highest possible accuracy. To achieve this goal, we will focus our attention on the following two research subjects: (i) the joint optimization of geometry and semantic classification from aerial images in a unified framework, and (ii) the exploitation of existing geographic information systems and web data to support these two sub-tasks. In addition, we will use web-based standard to efficiently represent the obtained results for fast modeling and data parsing.

The generation of realistic 3D models of whole cities has become a vibrant and highly competitive market through the recent activities of, most notably, Goggle Earth and Microsoft Virtual Earth. While the first generation of these systems only delivered high-quality zoomable images of the ground, the current trend is heavily geared towards 3D – that is, users can access three-dimensional height- fields of the terrain, and even 3D models of individual buildings. Simple building models, basically extruded polygons with different types of roofs, can be generated today from aerial images completely automatically. This is a solved problem. Far from solved, however, is the problem of generating automatically detailed buildings with façades. Input data for this problem are registered range maps obtained by stereo matching and sequences of highly overlapping thus redundant images (taken from a car driving in the road) where each pixel has not only a color but also a depth, a z-value. Although range maps can be directly rendered in principle, the data size is huge and, more importantly, the pixels have no semantics: A priori there is no difference between a pixel on the floor, on the wall, or on a door. But these shape semantics are required by all downstream applications using the city model. Shape grammars, on the other hand, have recently become (again) a popular method in research for representing 3D buildings. Their great advantage is that they allow to parameterize buildings, which can be used for populating virtual cities with believable architectural buildings, e.g., for 3D games. The goal of the CITYFIT project is, given highly redundant input imagery and range maps from an arbitrary building in Graz, to synthesize a shape grammar that, when evaluated, creates a clean, CAD- quality reconstruction of that building that fits the original data very closely and makes the semantics of all major architectural features explicit. These shape semantics can even be transferred back to inform the original data, so each of these “semantically enriched” data points can tell whether it belongs to ground, wall, or door.

The scanning electron microscope (SEM) is an important tool to examine very small structures. Its large magnification combined with good contrast and large depth of view make it possible to view and characterize microscopic structures in the sub-micron scale. In the recent years, the problem of dense surface reconstruction from multiple SEM images was a research topic on this institute. Reconstruction approaches like shape from stereo and shape from photometric stereo have been evaluated. This work presents a framework for automatic scene reconstruction from three images acquired by a scanning electron microscope. The basic assumption is that the specimen is tilted eucentrically in front of the camera, camera geometry is assumed to be unknown but constant over all views. It is shown that methods for estimating the trifocal tensor as well as modern auto-calibration approaches can be adapted to the imaging conditions in the SEM, and Euclidean scene structure can be retrieved from three uncalibrated views. The performance of the proposed framework is evaluated on synthetic data as well as real images. It is shown that Euclidean scene structure can be retrieved robustly under varying image noise and inaccurate initialization.