中国科学院自动化研究所机构知识库

Recently, visual understanding using unmanned aerial vehicles (UAVs) has gained significant attention due to its wide range of applications, including delivery, security investigation and surveillance. However, most existing UAV-based datasets only capture color images under ideal illumination and weather conditions, typically sunny days. This limitation fails to account for the complexity of real-world scenarios, such as cloudy or foggy weather, and nighttime conditions. Deep learning methods trained on color images with good lighting and weather conditions struggle to adapt to the complex visual scenes in these scenarios. Moreover, color images may not provide sufficient visual information under the complex visual scenes. To bridge this gap and meet the demands of real-world applications, we propose a large-scale RGB-Thermal Domain-incremental Object Detection (RTDOD) dataset in this paper. Our dataset includes RGB and thermal videos synchronously captured using calibrated color thermal cameras mounted on UAVs. It covers various weather conditions, from sunny to foggy to rainy, and spans from day to night. We sample and obtain approximately 16,200 pairs of images, and manually label dense annotations, including object bounding boxes and object categories. With the proposed dataset, we introduce a challenging domain-incremental object detection task. We also present a baseline approach that uses task-related gates to filter features for knowledge distillation to reduce forgetting. Experimental results on the RTDOD dataset demonstrate the effectiveness of our proposed method in domain-incremental object detection. To facilitate future research and development in domain-incremental object detection tasks on aerial images, the RTDOD dataset and our baseline model are made available at https://github.com/fenght96/RTDOD. ARTICLE INFO.