open_access
Refine
Document Type
- Article (6)
- Conference Proceeding (1)
Is part of the Bibliography
- yes (7)
Keywords
- ADAS (1)
- AI performance metrics (1)
- Artificial Intelligence (1)
- Computer Vision (1)
- DOAJ (1)
- Deep Neural Networks (1)
- Distribution Shift (1)
- IoU (1)
- MTF (1)
- Modulation Transfer Function (MTF) (1)
Department/institution
- Fachbereich - Elektro- & Informationstechnik (7) (remove)
Camera-based object detection is widely used in safety-critical applications such as advanced driver assistance systems (ADAS) and autonomous vehicle research. Road infrastructure has been designed for human vision, so computer vision, with RGB cameras, is a vital source of semantic information from the environment. Sensors, such as LIDAR and RADAR, are also often utilized for these applications; however, cameras provide a higher spatial resolution and color information. The spatial frequency response (SFR), or sharpness of a camera, utilized in object detection systems must be sufficient to allow a detection algorithm to localize objects in the environment over its lifetime reliably. This study explores the relationship between object detection performance and SFR. Six state-of-the-art object detection models are evaluated with varying levels of lens defocus. A novel raw image dataset is created and utilized, containing pedestrians and cars over a range of distances up to 100-m from the sensor. Object detection performance for each defocused dataset is analyzed over a range of distances to determine the minimum SFR necessary in each case. Results show that the relationship between object detection performance and lens blur is much more complex than previous studies have found due to lens field curvature, chromatic aberration, and astigmatisms. We have found that smaller objects are disproportionately impacted by lens blur, and different object detection models have differing levels of robustness to lens blur
Spatial precision and recall indices to assess the performance of instance segmentation algorithms
(2022)
Abstract The modulation-transfer function (MTF) is a fundamental optical metric to measure the optical quality of an imaging system. In the automotive industry it is used to qualify camera systems for ADAS/AD. Each modern ADAS/AD system includes evaluation algorithms for environment perception and decision making that are based on AI/ML methods and neural networks. The performance of these AI algorithms is measured by established metrics like Average Precision (AP) or precision-recall-curves. In this article we research the robustness of the link between the optical quality metric and the AI performance metric. A series of numerical experiments were performed with object detection and instance segmentation algorithms (cars, pedestrians) evaluated on image databases with varying optical quality. We demonstrate with these that for strong optical aberrations a distinct performance loss is apparent, but that for subtle optical quality differences – as might arise during production tolerances – this link does not exhibit a satisfactory correlation. This calls into question how reliable the current industry practice is where a produced camera is tested end-of-line (EOL) with the MTF, and fixed MTF thresholds are used to qualify the performance of the camera-under-test.
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.