Institut für Optische Systeme - IOS
Refine
Year of publication
Document Type
- Conference Proceeding (20)
- Article (9)
- Part of a Book (3)
- Doctoral Thesis (3)
- Master's Thesis (2)
- Bachelor Thesis (1)
- Book (1)
- Preprint (1)
- Report (1)
Keywords
Institute
Autonomous navigation on inland waters requires an accurate understanding of the environment in order to react to possible obstacles. Deep learning is a promising technique to detect obstacles robustly. However, supervised deep learning models require large data-sets to adjust their weights and to generalize to unseen data. Therefore, we equipped our research vessel with a laser scanner and a stereo camera to record a novel obstacle detection data-set for inland waters. We annotated 1974 stereo images and lidar point clouds with 3d bounding boxes. Furthermore, we provide an initial approach and a suitable metric to compare the results on the test data-set. The data-set is publicly available and seeks to make a contribution towards increasing the safety on inland waters.
Black-box variational inference (BBVI) is a technique to approximate the posterior of Bayesian models by optimization. Similar to MCMC, the user only needs to specify the model; then, the inference procedure is done automatically. In contrast to MCMC, BBVI scales to many observations, is faster for some applications, and can take advantage of highly optimized deep learning frameworks since it can be formulated as a minimization task. In the case of complex posteriors, however, other state-of-the-art BBVI approaches often yield unsatisfactory posterior approximations. This paper presents Bernstein flow variational inference (BF-VI), a robust and easy-to-use method flexible enough to approximate complex multivariate posteriors. BF-VI combines ideas from normalizing flows and Bernstein polynomial-based transformation models. In benchmark experiments, we compare BF-VI solutions with exact posteriors, MCMC solutions, and state-of-the-art BBVI methods, including normalizing flow-based BBVI. We show for low-dimensional models that BF-VI accurately approximates the true posterior; in higher-dimensional models, BF-VI compares favorably against other BBVI methods. Further, using BF-VI, we develop a Bayesian model for the semi-structured melanoma challenge data, combining a CNN model part for image data with an interpretable model part for tabular data, and demonstrate, for the first time, the use of BBVI in semi-structured models.
Incremental one-class learning using regularized null-space training for industrial defect detection
(2024)
One-class incremental learning is a special case of class-incremental learning, where only a single novel class is incrementally added to an existing classifier instead of multiple classes. This case is relevant in industrial defect detection scenarios, where novel defects usually appear during operation. Existing rolled-out classifiers must be updated incrementally in this scenario with only a few novel examples. In addition, it is often required that the base classifier must not be altered due to approval and warranty restrictions. While simple finetuning often gives the best performance across old and new classes, it comes with the drawback of potentially losing performance on the base classes (catastrophic forgetting [1]). Simple prototype approaches [2] work without changing existing weights and perform very well when the classes are well separated but fail dramatically when not. In theory, null-space training (NSCL) [3] should retain the basis classifier entirely, as parameter updates are restricted to the null space of the network with respect to existing classes. However, as we show, this technique promotes overfitting in the case of one-class incremental learning. In our experiments, we found that unconstrained weight growth in null space is the underlying issue, leading us to propose a regularization term (R-NSCL) that penalizes the magnitude of amplification. The regularization term is added to the standard classification loss and stabilizes null-space training in the one-class scenario by counteracting overfitting. We test the method’s capabilities on two industrial datasets, namely AITEX and MVTec, and compare the performance to state-of-the-art algorithms for class-incremental learning.
Particularly for manufactured products subject to aesthetic evaluation, the industrial manufacturing process must be monitored, and visual defects detected. For this purpose, more and more computer vision-integrated inspection systems are being used. In optical inspection based on cameras or range scanners, only a few examples are typically known before novel examples are inspected. Consequently, no large data set of non-defective and defective examples could be used to train a classifier, and methods that work with limited or weak supervision must be applied. For such scenarios, I propose new data-efficient machine learning approaches based on one-class learning that reduce the need for supervision in industrial computer vision tasks. The developed novelty detection model automatically extracts features from the input images and is trained only on available non-defective reference data. On top of the feature extractor, a one-class classifier based on recent developments in deep learning is placed. I evaluate the novelty detector in an industrial inspection scenario and state-of-the-art benchmarks from the machine learning community. In the second part of this work, the model gets improved by using a small number of novel defective examples, and hence, another source of supervision gets incorporated. The targeted real-world inspection unit is based on a camera array and a flashing light illumination, allowing inline capturing of multichannel images at a high rate. Optionally, the integration of range data, such as laser or Lidar signals, is possible by using the developed targetless data fusion method.
Using multi-camera matching techniques for 3d reconstruction there is usually the trade-off between the quality of the computed depth map and the speed of the computations. Whereas high quality matching methods take several seconds to several minutes to compute a depth map for one set of images, real-time methods achieve only low quality results. In this paper we present a multi-camera matching method that runs in real-time and yields high resolution depth maps. Our method is based on a novel multi-level combination of normalized cross correlation, deformed matching windows based on the multi-level depth map information, and sub-pixel precise disparity maps. The whole process is implemented completely on the GPU. With this approach we can process four 0.7 megapixel images in 129 milliseconds to a full resolution 3d depth map. Our technique is tailored for the recognition of non-technical shapes, because our target application is face recognition.
Das Projekt eFlow, an dem unter anderem die HTWG Konstanz seit 2012 forscht, simuliert mit Hilfe einer mathematischen Simulation wie sich Menschenmassen verhalten, wenn sie ein vorgegebenes Gelände verlassen sollen. Die Simulation baut auf einen Ansatz der Finite Elemente Methode auf, in der mehrere gekoppelte Differenzialgleichungen berechnet werden müssen. Diese Berechnungen erweisen sich gerade bei komplexen Szenarien mit großem Gelände und vielen Personen als sehr rechenintensiv. Ziel dieser Bachelorarbeit ist es ein Surrogate Modell zu erstellen, welches basierend auf machine-learning Ansätzen im spezifischen auf Regressionsmethoden Ergebnisse der Simulation vorhersagen soll. Somit müssen Datensätze generiert werden. Diese entstehen durch wiederholte Durchläufe der Simulation, in der jeweils die Eingabeparameter, die in das Regressionsmodell einfließen sollen variiert werden und mit dem entsprechenden Ergebnis der Simulation verknüpft werden. Die Regressionsansätze werden dabei pro Durchlauf komplexer, in dem jeweils zusätzliche Eingabeparameter mit in die Datengenerierung aufgenommen werden. Es soll überprüft werden, ob diese Simulation mittels machine-learning Ansätzen reproduzierbar ist. Basierend auf diesen Surrogate Modellen soll es möglich gemacht werden, Situationen in Echtzeit zu überprüfen, ohne dabei den Weg der rechenaufwendigen Simulation zu gehen. Die Ergebnisse bestätigen, dass die mathematische Simulation mittels Regression reproduzierbar ist. Es erweist sich jedoch als sehr rechenaufwendig, Daten zu sammeln, um genügend Eingabeparameter mit in die Regressionsmethode einfließen zu lassen. Diese Arbeit gestaltet somit eine Vorstudie zur Umsetzung eines ausgereiften Surrogate Modells, welches jegliche Eingabeparameter der Simulation berücksichtigen kann.
Random matrices are used to filter the center of gravity (CoG) and the covariance matrix of measurements. However, these quantities do not always correspond directly to the position and the extent of the object, e.g. when a lidar sensor is used.In this paper, we propose a Gaussian processes regression model (GPRM) to predict the position and extension of the object from the filtered CoG and covariance matrix of the measurements. Training data for the GPRM are generated by a sampling method and a virtual measurement model (VMM). The VMM is a function that generates artificial measurements using ray tracing and allows us to obtain the CoG and covariance matrix that any object would cause. This enables the GPRM to be trained without real data but still be applied to real data due to the precise modeling in the VMM. The results show an accurate extension estimation as long as the reality behaves like the modeling and e.g. lidar measurements only occur on the side facing the sensor.
Motion estimation is an essential element for autonomous vessels. It is used e.g. for lidar motion compensation as well as mapping and detection tasks in a maritime environment. Because the use of gyroscopes is not reliable and a high performance inertial measurement unit is quite expensive, we present an approach for visual pitch and roll estimation that utilizes a convolutional neural network for water segmentation, a stereo system for reconstruction and simple geometry to estimate pitch and roll. The algorithm is validated on a novel, publicly available dataset recorded at Lake Constance. Our experiments show that the pitch and roll estimator provides accurate results in comparison to an Xsens IMU sensor. We can further improve the pitch and roll estimation by sensor fusion with a gyroscope. The algorithm is available in its implementation as a ROS node.
Targetless Lidar-camera registration is a repeating task in many computer vision and robotics applications and requires computing the extrinsic pose of a point cloud with respect to a camera or vice-versa. Existing methods based on learning or optimization lack either generalization capabilities or accuracy. Here, we propose a combination of pre-training and optimization using a neural network-based mutual information estimation technique (MINE [1]). This construction allows back-propagating the gradient to the calibration parameters and enables stochastic gradient descent. To ensure orthogonality constraints with respect to the rotation matrix we incorporate Lie-group techniques. Furthermore, instead of optimizing on entire images, we operate on local patches that are extracted from the temporally synchronized projected Lidar points and camera frames. Our experiments show that this technique not only improves over existing techniques in terms of accuracy, but also shows considerable generalization capabilities towards new Lidar-camera configurations.
Image novelty detection is a repeating task in computer vision and describes the detection of anomalous images based on a training dataset consisting solely of normal reference data. It has been found that, in particular, neural networks are well-suited for the task. Our approach first transforms the training and test images into ensembles of patches, which enables the assessment of mean-shifts between normal data and outliers. As mean-shifts are only detectable when the outlier ensemble and inlier distribution are spatially separate from each other, a rich feature space, such as a pre-trained neural network, needs to be chosen to represent the extracted patches. For mean-shift estimation, the Hotelling T2 test is used. The size of the patches turned out to be a crucial hyperparameter that needs additional domain knowledge about the spatial size of the expected anomalies (local vs. global). This also affects model selection and the chosen feature space, as commonly used Convolutional Neural Networks or Vision Image Transformers have very different receptive field sizes. To showcase the state-of-the-art capabilities of our approach, we compare results with classical and deep learning methods on the popular dataset CIFAR-10, and demonstrate its real-world applicability in a large-scale industrial inspection scenario using the MVTec dataset. Because of the inexpensive design, our method can be implemented by a single additional 2D-convolution and pooling layer and allows particularly fast prediction times while being very data-efficient.
