SOLAS 1.1: Automotive Optical Simulation in Computer Vision

Automotive datasets are typically captured using a small number of cameras, with each camera fixed at a single focus setting. In practice, however, camera modules exhibit unit-to-unit variability in their effective focus due to manufacturing tolerances. Since perception models are usually trained on images captured at one nominal focus position, real-world deviations in focus can introduce a domain mismatch that degrades perception performance. We demonstrate this effect by simulating two different optical systems on synthetic and real images with fields of view of <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$100^\circ$</tex-math></inline-formula> and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$150^\circ$</tex-math></inline-formula>. For all simulations, we utilise the Python-based ray-tracing library KrakenOS, an open-source optical simulation tool. By assigning each optical system to a suitable dataset, we degrade the held-out test data of four public automotive datasets: KITTI, Virtual KITTI 2.0, Woodscape, and Parallel Domain Woodscape. We evaluate the impact of applying optical defocus on 2D Object Detection models with the popular OpenMMLab toolkit for MMDetection and the YOLOv11 architecture. For each optical system, we simulate 9 defocus settings on the test data, representative of the production tolerance range for camera defocus. The results show that object detection performance degrades as the magnitude of defocus increases. Align DETR, despite having the second fewest parameters, establishes the strongest baseline and remains robust under modest defocus (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$|\Delta z|\le 20,\mu \mathrm{m}$</tex-math></inline-formula>) across all datasets. However, at extreme defocus (<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\pm 100,\mu \mathrm{m}$</tex-math></inline-formula>), YOLOv11x surpasses Align DETR by 1.5%-12.2% mAP<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$_{50:95}$</tex-math></inline-formula> across all datasets. Finally, we show that defocus-augmented training of Align DETR, recovers the performance drop caused by the defocus in the held-out test data.