Dietrich Paulus

Automated segmentation of the optic nerve head for diagnosis of glaucoma

Glaucoma is the second most common cause of blindness worldwide. Low awareness and high costs connected to glaucoma are reasons to improve methods of screening and therapy. A well-established method for diagnosis of glaucoma is the examination of the optic nerve head using scanning-laser-tomography. This system acquires and analyzes the surface topography of the optic nerve head. The analysis that leads to a diagnosis of the disease depends on prior manual outlining of the optic nerve head by an experienced ophthalmologist. Our contribution presents a method for optic nerve head segmentation and its validation. The method is based on morphological operations, Hough transform, and an anchored active contour model. The results were validated by comparing the performance of different classifiers on data from a case-control study with contours of the optic nerve head manually outlined by an experienced ophthalmologist. We achieved the following results with respect to glaucoma diagnosis: linear discriminant analysis with 27.7% estimated error rate for automated segmentation (aut) and 26.8% estimated error rate for manual segmentation (man), classification trees with 25.2% (aut) and 22.0% (man) and bootstrap aggregation with 22.2% (aut) and 13.4% (man). It could thus be shown that our approach is suitable for automated diagnosis and screening of glaucoma.

Texture-Based Polyp Detection in Colonoscopy

Colonoscopy is one of the best methods for screening colon cancer. A variety of research groups have proposed methods for automatic detection of polyps in colonoscopic images to support the doctors during examination. However, the problem can still not be assumed as solved. The major drawback of many approaches is the amount and quality of images used for classifier training and evaluation. Our database consists of more than four hours of high resolution video from colonoscopies which were examined and labeled by medical experts. We applied four methods of texture feature extraction based on Grey-Level-Co-occurence and Local-Binary-Patterns. Using this data, we achieved classification results with an area under the ROC-curve of up to 0.96.

Study of parameterizations for the rigid body transformations of the scan registration problem

The iterative closest point (ICP) algorithm is the de facto standard for geometric alignment of three-dimensional models when an initial relative pose estimate is available. The basis of the algorithm is the minimization of an error function that takes point correspondences into account. Four closed-form solution methods are known for minimizing this function. This paper presents novel linear solutions to the scan registration problem, i.e., to the problem of putting and aligning 3D scans in a common coordinate system. We extend the methods for registering n-scans in a global and simultaneous fashion, such that the registration of the nth scan influences all previous registrations in one step.

Three-dimensional computer vision for tooth restoration

If a person with carious lesions needs or requests crowns or inlays, these dental fillings have to be manufactured for each tooth and each person individually. We survey computer vision techniques which can be used to automate this process. We introduce three particular applications which are concerned with the reconstruction of surface information. The first one aims at building up a database of normalized depth images of posterior teeth and at extracting characteristic features from these images. In the second application, a given occlusal surface of a posterior tooth with a prepared cavity is digitally reconstructed using an intact model tooth from a given database. The calculated surface data can then be used for automatic milling of a dental prosthesis, e.g. from a preshaped ceramic block. In the third application a hand-made provisoric wax inlay or crown can be digitally scanned by a laser sensor and copied three dimensionally into a different material such as ceramic. The results are converted to a format required by the computer-integrated manufacturing (CIM) system for automatic milling.

Highlight substitution in light fields

Highlights occur especially when recording medical (color) images during micro-invasive operations. They disturb the physicians who can sometimes only guess the tissue at the position of the highlights. We present a new technique of highlight removal. A so-called light field is generated from the recorded image sequence. Then a binary highlight mask is computed for each image and used as confidence map for the light field pixels. The result is a light field in which pixels at highlight positions are interpolated by pixels which were not over-imposed by highlights. This leads to light fields with better images. We demonstrate and evaluate the technique on medical and synthetic image sequences.

An evaluation of open source SURF implementations

SURF (Speeded Up Robust Features) is a detector and descriptor of local scale- and rotation-invariant image features. By using integral images for image convolutions it is faster to compute than other state-of-the-art algorithms, yet produces comparable or even better results by means of repeatability, distinctiveness and robustness. A library implementing SURF is provided by the authors. However, it is closedsource and thus not suited as a basis for further research. Several open source implementations of the algorithm exist, yet it is unclear how well they realize the original algorithm. We have evaluated different SURF implementations written in C++ and compared the results to the original implementation. We have found that some implementations produce up to 33% lower repeatability and up to 44% lower maximum recall than the original implementation, while the implementation provided with the software Pan-o-matic produced almost identical results. We have extended the Pan-o-matic implementation to use multithreading, resulting in an up to 5.1 times faster computation on an 8-core machine. We describe our comparison criteria and our ideas that lead to the speed-up. Our software is put into the public domain.

2D/3D image registration on the GPU

We present a method that performs a rigid 2D/3D image registration efficiently on the Graphical Processing Unit (GPU). As one main contribution of this paper, we propose an efficient method for generating realistic DRRs that are visually similar to x-ray images. Therefore, we model some of the electronic post-processes of current x-ray C-arm-systems. As another main contribution, the GPU is used to compute eight intensity-based similarity measures between the DRR and the x-ray image in parallel. A combination of these eight similarity measures is used as a new similarity measure for the optimization. We evaluated the performance and the precision of our 2D/3D image registration algorithm using two phantom models. Compared to a CPU + GPU algorithm, which calculates the similarity measures on the CPU, our GPU algorithm is between three and six times faster. In contrast to single similarity measures, our new similarity measure achieved precise and robust registration results for both phantom models.

Terrain drivability analysis in 3D laser range data for autonomous robot navigation in unstructured environments

Three-dimensional laser range finders provide autonomous systems with vast amounts of information. However, autonomous robots navigating in unstructured environments are usually not interested in every geometric detail of their surroundings. Instead, they require real-time information about the location of obstacles and the condition of drivable areas.In this paper, we first present grid-based algorithms for classifying regions as either drivable or not. In a subsequent step, drivable regions are further examined using a novel algorithm which determines the local terrain roughness. This information can be used by a path planning algorithm to decide whether to prefer a rough, muddy area, or a plain street, which would not be possible using binary drivability information only.

Probabilistic terrain classification in unstructured environments

Autonomous navigation in unstructured environments is a complex task and an active area of research in mobile robotics. Unlike urban areas with lanes, road signs, and maps, the environment around our robot is unknown and unstructured. Such an environment requires careful examination as it is random, continuous, and the number of perceptions and possible actions are infinite. We describe a terrain classification approach for our autonomous robot based on Markov Random Fields (MRFs ) on fused 3D laser and camera image data. Our primary data structure is a 2D grid whose cells carry information extracted from sensor readings. All cells within the grid are classified and their surface is analyzed in regard to negotiability for wheeled robots. Knowledge of our robots egomotion allows fusion of previous classification results with current sensor data in order to fill data gaps and regions outside the visibility of the sensors. We estimate egomotion by integrating information of an IMU, GPS measurements, and wheel odometry in an extended Kalman filter. In our experiments we achieve a recall ratio of about 90% for detecting streets and obstacles. We show that our approach is fast enough to be used on autonomous mobile robots in real time.

Real-time 3D mapping of rough terrain: A field report from Disaster City

Mobile systems for mapping and terrain classification are often tested on datasets of intact environments only. The behavior of the algorithms in unstructured environments is mostly unknown. In safety, security and rescue environments, the robots have to handle much rougher terrain. Therefore, there is a need for 3D test data that also contains disaster scenarios. During the Response Robot Evaluation Exercise in March 2010 in Disaster City, College Station, Texas (USA), a comprehensive dataset was recorded containing the data of a 3D laser range finder, a GPS receiver, an IMU and a color camera. We tested our algorithms (for terrain classification and 3D mapping) with the dataset, and will make the data available to give other researchers the chance to do the same. We believe that this captured data of this well known location provides a valuable dataset for the USAR robotics community, increasing chances of getting more comparable results.