Bridging the Scale Gap: Balanced Tiny and General Object Detection in Remote Sensing Imagery

Zhi-Cheng Zhao, Yin Huang, Lingma Sun, Chenglong Li, Jin TangarXiv 2025 · DOI · arXiv

Abstract

Remote sensing imagery presents a unique challenge for object detection due to the extreme "scale gap" between tiny objects, such as vehicles, and massive structures like industrial complexes. Traditional detectors often specialize in one scale at the expense of the other. This paper introduces ScaleBridge-Det, a novel large-scale detection framework that utilizes a scale-adaptive Mixture-of-Experts (MoE) architecture and density-guided dynamic queries. By balancing representations across scales and dynamically adjusting computational resources based on scene density, ScaleBridge-Det achieves state-of-the-art performance on tiny object benchmarks (35.7% mAP on AI-TOD-v2) while maintaining robustness across diverse remote sensing domains.

Introduction: The Scale Competition Problem

In satellite and drone-captured imagery, a single scene often contains objects spanning several orders of magnitude in size. This "scale competition" leads to two common failure modes in existing detectors: specialization in tiny objects that loses global context, or generalist models where large-scale features overwhelm signals from smaller targets. Furthermore, varying object densities—from sparse rural areas to crowded urban centers—render fixed-query Transformer architectures inefficient.

Visualizing the scale gap and density variations across different remote sensing scenarios, including airports, urban intersections, and parking lots.

Methodology: ScaleBridge-Det

ScaleBridge-Det addresses these issues through two core innovations:

1. Routing-Enhanced Mixture Attention (REM)

This module employs a Mixture-of-Experts (MoE) approach, integrating multiple backbones (e.g., ResNet, ViT, Swin) specialized for different scales. An adaptive routing mechanism dynamically assigns weights to experts. To prevent "expert collapse," a category-wise top-1 selection ensures that at least one expert for each scale category (tiny, general, mixed) is always active, preserving fine-grained details even in complex scenes.

2. Density-Guided Dynamic Query (DGQ)

Instead of a static number of object queries, DGQ predicts an object density map of the scene. Based on this map, the model allocates queries across four tiers (ranging from 900 for sparse scenes to 2,000 for extremely dense ones). It also uses spatial priors from the density map to initialize query positions in high-probability regions, preventing query starvation in crowded areas.

Experimental Results

The framework was evaluated on AI-TOD-v2, DTOD, and VisDrone datasets. ScaleBridge-Det demonstrated superior performance, particularly in dense environments where traditional models struggle.

Detection performance in a complex urban intersection, showcasing the model's ability to handle high density and varied scales simultaneously.

On the AI-TOD-v2 benchmark, the model achieved 35.7% mAP, significantly outperforming general-purpose foundation models like DINOv3 (7B parameters). In the dense DTOD dataset, it reached 29.8% AP50, a 5.6 percentage point improvement over previous state-of-the-art methods. Cross-domain evaluation on VisDrone further confirmed the model's robustness, achieving 30.7% mAP without any fine-tuning on drone-specific data.

Conclusion

ScaleBridge-Det successfully bridges the gap between tiny and general object detection in remote sensing. By incorporating architectural priors for scale and density, it provides a balanced and efficient solution for complex aerial imagery. Future work will focus on developing lightweight versions for real-time edge deployment on UAVs and extending the framework to other modalities such as Synthetic Aperture Radar (SAR).

Summary updated

BibTeX
@misc{zhao2025bridging,
  author = {Zhi-Cheng Zhao and Yin Huang and Lingma Sun and Chenglong Li and Jin Tang},
  title = {{Bridging the Scale Gap: Balanced Tiny and General Object Detection in Remote Sensing Imagery}},
  year = {2025},
  eprint = {2512.01665},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.01665},
}
Figure