Integritets filtrering i bilder : en jämförande studie av registreringsskyltsdetektering i höguplösta ekvirektangulära panoraman
Jonatan Tuvstedt · KTH Publication Database DiVA (KTH Royal Institute of Technology) · 2026
Object detection is a crucial step for automatically filtering out personal information from images, a task that is becoming ever more important as regulations around data protection tightens. One area where object detection faces significant challenges is in non-planar images like equirectangular panoramas, an image type becoming increasingly common in traffic environments to represent 360 degree fields of view. This thesis investigates how four state of the art object detection models YOLOX, RF-DETR, RT-DETRv2 and YOLOv11 compares in terms of accuracy and computational performance for licence plate detection on large equirectangular panoramas in traffic environments. This comparison is conducted on both raw panoramas and re-projections of the panoramas to a rectilinear perspective domain, and contrasted with the performance on a comparable planar dataset from the same source. Additionally to adapt to the very high resolutions of the panoramas, images are broken down into patches before inference using a sliding window approach, and the early stopping model MTC, originally developed for remote sensing, is investigated as a means of limiting the number of full inference runs required. Of the four models YOLOX consistently performs the best. Furthermore, the rectified version of the equirectangular panoramas outperformed the raw panoramas, significantly so for YOLOv11 and RF-DETRv2, at a major computational cost. Finally MTC has only minimal effects on final detection accuracies while ending between 2/3s and 4/5s of all inference runs early, though its computational performance on GPUs is poor meaning additional work is needed for it to be a viable option. These results demonstrate that state of the art object detection models can effectively be adapted for equirectangular panoramas, and that rectification can be an effective method to improve detection performance, but that the effectiveness depends on the model. Finally MTC is shown to be a feasible technique for limiting the number of inference runs in very large sparse images at a minimal accuracy cost, but the exact implementation needs additional optimisation work.