"MI-OAD Dataset"

"MI-OAD: Multi-instance Open-set Aerial DatasetMI-OAD is a large-scale benchmark dataset designed for language-guided open-set aerial object detection. It is constructed using the OS-W2S Label Engine \u2014 an automatic annotation pipeline that integrates an open-source vision-language model with image-operation-based preprocessing and BERT-based postprocessing. The dataset contains 163,023 aerial images and over 2 million image-caption pairs with multi-granularity annotations at word, phrase, and sentence levels, covering 100 object categories across diverse aerial scenes captured from various altitudes and platforms. Unlike existing datasets, MI-OAD supports flexible multi-instance retrieval per caption, enabling more realistic grounding scenarios. It is approximately 40\u00d7 larger than existing remote sensing visual grounding datasets. Training on MI-OAD improves Grounding DINO by +31.1 AP50 under zero-shot transfer, establishing new state-of-the-art performance on aerial object detection and remote sensing visual grounding benchmarks."