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

Remote sensing object detection has historically been fragmented, with models often specialized for specific sensors, resolutions, or narrow category sets. This paper addresses these limitations by introducing LEVIRDet-159, a comprehensive dataset, and LEVIRDetNet, a foundation model designed to handle the scale, density, and semantic hierarchy inherent in remote sensing imagery.

LEVIRDet-159 is currently the largest remote sensing object detection dataset, featuring 174,488 images, 2.56 million bounding boxes, and 700k fine-grained annotations. It organizes data under a multi-level taxonomy, covering 30 parent categories and 159 specific types. Unlike previous datasets that often focus on a single object family or resolution, LEVIRDet-159 integrates diverse sources—including satellite, aerial, and map-service imagery—to provide a broad range of object sizes and scene densities. The authors employed a rigorous data engine to standardize heterogeneous annotations into a unified tight horizontal bounding box (HBB) protocol, ensuring consistency across the entire collection.

LEVIRDetNet is a scale-hierarchy-aware foundation model built to leverage the diversity of the LEVIRDet-159 dataset. It incorporates three key innovations:

• Online GSD Prediction: Estimates the Ground Sampling Distance to provide physical scale cues for query modulation.
• Dynamic Query Allocation: Adjusts the number of object hypotheses based on the density of the scene.
• Hierarchy-Aware Detection Head: Enables the model to learn from mixed-granularity labels (e.g., distinguishing between a general 'aircraft' and a specific 'Boeing 747') while remaining compatible with standard, flat-category evaluation metrics.

In a zero-shot, target-training-free evaluation, LEVIRDetNet outperformed existing state-of-the-art methods on 9 external benchmarks. It achieved an average improvement of 5.02 mAP over fully supervised competing models. These results suggest that by explicitly modeling the physical and semantic properties of remote sensing data, it is possible to build a foundation model that generalizes effectively across diverse sensors and environments without requiring domain-specific fine-tuning.