Refine
Year of publication
Document Type
- Article (22)
- Conference Proceeding (7)
- Preprint (2)
- Diploma Thesis (1)
- Doctoral Thesis (1)
- Patent (1)
Keywords
- Artificial Intelligence (3)
- DEAL (2)
- DFG Publikationskosten (2)
- HSD Publikationsfonds (2)
- Machine Learning (2)
- ADAS (1)
- AI performance metrics (1)
- Computer Vision (1)
- Computers and Society (1)
- DOAJ (1)
Department/institution
- Fachbereich - Elektro- & Informationstechnik (34) (remove)
We present a novel metric Spatial Recall Index to assess the performance of machine-learning (ML) algorithms for automotive applications, focusing on where in the image which performance occurs. Typical metrics like intersection-over-union (IoU), precisionrecallcurves or average precision (AP) quantify the performance over a whole database of images, neglecting spatial performance variations. But as the optics of camera systems are spatially variable over the field of view, the performance of ML-based algorithms is also a function of space, which we show in simulation: A realistic objective lens based on a Cooke-triplet that exhibits typical optical aberrations like astigmatism and chromatic aberration, all variable over field, is modeled. The model is then applied to a subset of the BDD100k dataset with spatially-varying kernels. We then quantify local changes in the performance of the pre-trained Mask R-CNN algorithm. Our examples demonstrate the spatial dependence of the performance of ML-based algorithms from the optical quality over field, highlighting the need to take the spatial dimension into account when training ML-based algorithms, especially when looking forward to autonomous driving applications.
Spatial precision and recall indices to assess the performance of instance segmentation algorithms
(2022)
We present an experimental technique that allows the simultaneous spatial imaging and spectral analysis of falling droplets that exhibit lasing. Single droplet investigations serve as, among other purposes, a preliminary study for spray and combustion researchers. The described setup provides a valuable tool for the evaluation of microdroplet investigations with laser-spectroscopic techniques that rely on laser-induced fluorescence (LIF) or similar spectroscopical phenomena. The emphasis is that both spatial and spectral information are obtained from single-shot images of a falling droplet. Furthermore, combining spatial imaging and a spatially resolving optical multichannel analyzer makes a pointwise rastering of the droplets spectrum possible. This allows for the (almost) unambiguous determination of sources of influence on the spectrum of these droplets-such as geometrical distortion and lasing, nondissolved tracer lumps, and similar phenomena. Although the focus is on the experimental technique itself, we supplement detailed studies of lasing in falling microdroplets. These results were obtained with the aim of developing a system for measuring temperature distributions in droplets and sprays. In the light of these results the practice of calibrating a droplets spectrum by use of a bulk liquid sample needs to be critically reviewed.
We had already demonstrated a numerical model for the point spread function (PSF) of an optical system that can efficiently model both the experimental measurements and the lens design simulations of the PSF. The novelty lies in the portability and the parameterization of this model, which allow for completely new ways to validate optical systems, which is especially interesting not only for mass production optics such as in the automotive industry but also for ophthalmology. The numerical basis for this model is a nonlinear regression of the PSF with an artificial neural network (ANN). After briefly describing both the principle and the applications of the model, we then discuss two optically important aspects: the spatial resolution and the accuracy of the model. Using mean squared error (MSE) as a metric, we vary the topology of the neural network, both in the number of neurons and in the number of hidden layers. Measurement and simulation of a PSF can have a much higher spatial resolution than the typical pixel size used in current camera sensors. We discuss the influence this has on the topology of the ANN. The relative accuracy of the averaged pixel MSE is below 10 − 4, thus giving confidence that the regression does indeed model the measurement data with good accuracy. This article is only the starting point, and we propose several research avenues for future work.
In a previous work we have demonstrated a novel numerical model for the point spread function (PSF) of an optical system that can efficiently model both experimental measurements and lens design simulations of the PSF. The novelty lies in the portability and the parameterization of this model, which allows for completely new ways to validate optical systems, which is especially interesting for mass production optics like in the automotive industry, but also for ophtalmology. The numerical basis for this model is a non-linear regression of the PSF with an artificial neural network (ANN). In this work we examine two important aspects of this model: the spatial resolution and the accuracy of the model. Measurement and simulation of a PSF can have a much higher resolution then the typical pixel size used in current camera sensors, especially those for the automotive industry. We discuss the influence this has on on the topology of the ANN and the final application where the modeled PSF is actually used. Another important influence on the accuracy of the trained ANN is the error metric which is used during training. The PSF is a distinctly non-linear function, which varies strongly over field and defocus, but nonetheless exhibits strong symmetries and spatial relations. Therefore we examine different distance and similarity measures and discuss its influence on the modeling performance of the ANN.
The importance of high data quality is increasing with the growing impact and distribution of ML systems and big data. Also, the planned AI Act from the European commission defines challenging legal requirements for data quality especially for the market introduction of safety relevant ML systems. In this paper, we introduce a novel approach that supports the data quality assurance process of multiple data quality aspects. This approach enables the verification of quantitative data quality requirements. The concept and benefits are introduced and explained on small example data sets. How the method is applied is demonstrated on the well-known MNIST data set based an handwritten digits.