开放词汇检测器在航空影像上的迁移性：一项比较评估

Open-vocabulary object detection (OVD) leverages vision-language models for zero-shot recognition of novel categories, achieving strong performance on natural images. Nevertheless, its transferability to aerial imagery remains largely unexplored. This research presents the first systematic benchmark evaluating five state-of-the-art OVD models on the LAE-80C aerial dataset, which comprises 3,592 images and 80 categories, under strict zero-shot conditions. The experimental protocol is meticulously designed to isolate the models' understanding of semantic concepts from their generalization capabilities to specific image features. An in-depth analysis of these models' performance reveals the challenges and opportunities for current open-vocabulary detectors in the aerial imagery domain. Results indicate that despite their excellent performance on natural images, a significant performance drop is observed on aerial imagery due to differences in perspective, object scale, background complexity, and category distribution. Specifically, the zero-shot recognition capabilities of these models struggle with detecting small, densely packed, or highly occluded aerial targets. The evaluation also investigates the impact of different vision-language model architectures on performance, identifying which model design characteristics contribute to improved detection accuracy and robustness in aerial imagery scenarios. The experimental data provides a crucial benchmark and direction for future development of more effective open-vocabulary detectors tailored for aerial imagery, emphasizing the need for further research to bridge the domain gap between natural and aerial images for more efficient zero-shot object recognition.