Real-world applications of machine learning models are often subject to legal or policy-based regulations. Some of these regulations require ensuring the validity of the model, i.e., the approximation error being smaller than a threshold. A global metric is generally too insensitive to determine the validity of a specific prediction, whereas evaluating local validity is costly since it requires gathering additional data. We propose learning the model error to acquire a local validity estimate while reducing the amount of required data through active learning. Using model validation benchmarks, we provide empirical evidence that the proposed method can lead to an error model with sufficient discriminative properties using a relatively small amount of data. Furthermore, an increased sensitivity to local changes of the validity bounds compared to alternative approaches is demonstrated.

Modern neural networks have found to be miscalibrated in terms of confidence calibration, i.e., their predicted confidence scores do not reflect the observed accuracy or precision. Recent work has introduced methods for post-hoc confidence calibration for classification as well as for object detection to address this issue. Especially in safety critical applications, it is crucial to obtain a reliable self-assessment of a model. But what if the calibration method itself is uncertain, e.g., due to an insufficient knowledge base? We introduce Bayesian confidence calibration - a framework to obtain calibrated confidence estimates in conjunction with an uncertainty of the calibration method. Commonly, Bayesian neural networks (BNN) are used to indicate a network's uncertainty about a certain prediction. BNNs are interpreted as neural networks that use distributions instead of weights for inference. We transfer this idea of using distributions to confidence calibration. For this purpose, we use stochastic variational inference to build a calibration mapping that outputs a probability distribution rather than a single calibrated estimate. Using this approach, we achieve state-of-the-art calibration performance for object detection calibration. Finally, we show that this additional type of uncertainty can be used as a sufficient criterion for covariate shift detection. All code is open source and available at https://github.com/EFS-OpenSource/calibration-framework.

In current state-of-the-art systems for object detection and classification a huge amount of data is needed. Even if large databases are available, some classes are typically underrepresented and therefore the classifier is not able to capture the variability in appearance. In this work we present a novel method to enrich the training database with natural looking synthetic images. The method can be used to transfer the object appearance from one image (template image) to another image (base image) containing a different object of the same or a similar category. In order to preserve natural appearance and avoid artifacts we only use the gray-level values of the base image for synthesis. The main contribution of this work is an extension of the shift-map approach [1]. An appropriate optimization criteria for the used Markov Random Field (MRF) is defined and the MRF is embedded into a general framework for training data synthesis, which is exemplary tailored to the generation of traffic signs. The influence of using synthetic images is evaluated using a convolutional neural network (CNN).

In this paper a new crosswalk detection strategy is presented. The application area is narrowed down to driver assistance systems to guarantee a reliable detection result and to benefit from the properties of a vehicle mounted camera. The main contribution of this paper is an efficient method to detect line segments that are typical for road markings like lanes and crosswalks. Therefore an adaptive Haar-like filter is applied by means of an integral image, where the pass-band of the filter is adjusted according to the distance from the vehicle. The detected segments are then combined to regions of interest by means of prior knowledge of crosswalk dimensions. These regions are then passed to a simple classifying module that utilizes a set of moment invariants for classification. The final detection result is then tracked by a standard Kalman filter.

Although recent statistics demonstrate a decrease in pedestrian fatalities, the absolute number of accident related deaths is sufficiently high to justify research in the area of vulnerable road user protection. Research has shown that situation awareness, which requires a significant amount of context information, is critical to timely intervention. Altered pedestrian behavior, due to traffic regulation, requires context information. For example, crosswalk presence and location knowledge can be of importance in pedestrian crossing scenarios. Therefore this paper discusses the implementation of a crosswalk detection algorithm, using Fourier transformation, augmented bipolarity, inverse perspective mapping template matching and edge orientation ratios for classification.