LEVIRDet: A Million-Scale 159-Category Dataset and Foundation Model for Universal Remote Sensing Object Detection

Researchers have introduced LEVIRDet-159, a remote sensing object detection dataset containing 159 categories, 2.56 million bounding boxes, and 700,000 fine-grained annotations, making it the largest and most comprehensive benchmark of its kind, according to a paper posted to arXiv on June 24 [1]. The dataset exceeds the largest existing remote sensing object detection dataset in every key scale dimension, holding approximately seven times more images, six times more object instances, and four times more categories [1]. It is organized under a multi-level taxonomy, an effort to address fragmentation in a field where most benchmarks focus on limited categories, fixed spatial resolutions, or a single sensor [1]. Alongside the dataset, the authors designed LEVIRDetNet, a scale-hierarchy-aware detection foundation model for universal remote sensing object detection [1]. The model couples online visual Ground Sampling Distance prediction, GSD-conditioned query modulation and allocation, and a hierarchy-aware detection head for mixed-granularity supervision [1]. Under stringent evaluation settings, LEVIRDetNet demonstrated strong cross-domain generalization, achieving state-of-the-art detection performance on nine external benchmarks without target-domain training or fine-tuning [1]. It improved the strongest fully supervised competing methods by an average of 5.02 mAP under each benchmark's primary metric [1]. The paper appears on arXiv, a preprint server that has become a central distribution channel for machine learning research. Platforms such as Hugging Face have built infrastructure to link arXiv papers with models, datasets, and interactive demos, allowing users to find artifacts related to a paper and enabling community discussion [4]. Hugging Face and arXiv have also collaborated to embed demos directly alongside papers on arXiv abstract pages, so readers can try state-of-the-art research without writing code [5]. The release of large-scale open benchmarks has been a recurring pattern in AI development. In the language domain, companies such as DeepSeek have released open-weight models that rival proprietary systems at lower reported training costs, a move that sent what observers called "shock waves" through the industry [7]. Alibaba Cloud's Qwen family of large language models is distributed under open-source licenses including Apache 2.0 [9]. The LEVIRDet-159 dataset and trained models will be released accompanying the final paper, the authors stated [1].