The EVD4UAV dataset is proposed as a benchmark for evading vehicle detection in UAV imagery by introducing diverse altitudes, vehicle attributes, and fine-grained annotations. The study explores white-box and black-box attack methods on classic deep neural network-based object detectors, revealing the challenges and implications of altitude-sensitive attacks.
The content discusses the limitations of existing public UAV datasets in addressing adversarial patch-based vehicle detection attacks due to overlooking altitude variations and detailed annotations. The proposed EVD4UAV dataset includes 6,284 images with 90,886 annotated vehicles at different altitudes (50m, 70m, 90m) and clear top-view images for effective attack studies.
Experimental results show that representative attack methods struggle to achieve robust altitude-insensitive performance on EVD4UAV. White-box and black-box attack scenarios are explored using different object detectors like Faster R-CNN, DETR, and YOLOv8. The study emphasizes the need for further research on altitude-insensitive attack strategies using the EVD4UAV dataset.
Para outro idioma
do conteúdo fonte
arxiv.org
Principais Insights Extraídos De
by Huiming Sun,... às arxiv.org 03-11-2024
https://arxiv.org/pdf/2403.05422.pdfPerguntas Mais Profundas