Object detection is one of the most crucial applications of computer vision, with applications in numerous fields, ranging from autonomous robots to virtual and mixed reality systems. Consequently, the reliability and robustness of these methods is various uncontrolled settings is paramount. However, the performance of object detection may be significantly hindered by interfering circumstances, such as environmental effects (fog, rain, etc.) or setup-induced problems, such as occlusion. This work is proposing a novel method for increasing the robustness of existing object detection methods to full or partial occlusion by combining multiple views of the same scene. Our method introduces the geometric framework used in Neural Radiance Fields (NeRFs) to provide a solution that can be easily integrated into any existing object detector, and requires no on-line training. Our evaluation on our test set demonstrates that our method significantly outperforms the state-of-the-art YOLO v7.
NeRF-YOLO: Detecting Occluded Objects via Multi-view Geometric Aggregation
Mechan. Machine Science
International Symposium on Measurements and Control in Robotics ; 2023 ; Iasi, Romania September 21, 2023 - September 22, 2023
25th International Symposium on Measurements and Control in Robotics ; Kapitel : 2 ; 13-22
02.04.2024
10 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Detecting potentially occluded objects for autonomous vehicles
Europäisches Patentamt | 2025
|DETECTING POTENTIALLY OCCLUDED OBJECTS FOR AUTONOMOUS VEHICLES
Europäisches Patentamt | 2021
|Detecting potentially occluded objects for autonomous vehicles
Europäisches Patentamt | 2022
|