Learning Where to Focus: Density-Driven Guidance for Detecting Dense Tiny Objects

TL;DR

This paper tackles the problem of detecting densely packed tiny objects in high-resolution remote sensing imagery. It introduces DRMNet, a density-driven framework that uses a Density Generation Branch to create spatial priors and two modules, Dense Area Focusing Module and Dual Filter Fusion Module, to enable adaptive region-aware learning and density-guided frequency fusion. The approach yields state-of-the-art performance on AI-TOD and DTOD, particularly boosting detection in highly dense and occluded scenarios. By integrating explicit density priors into both spatial attention and frequency-domain fusion, the method achieves improved feature discrimination while maintaining computational efficiency, with practical impact on remote sensing analysis tasks.

Abstract

High-resolution remote sensing imagery increasingly contains dense clusters of tiny objects, the detection of which is extremely challenging due to severe mutual occlusion and limited pixel footprints. Existing detection methods typically allocate computational resources uniformly, failing to adaptively focus on these density-concentrated regions, which hinders feature learning effectiveness. To address these limitations, we propose the Dense Region Mining Network (DRMNet), which leverages density maps as explicit spatial priors to guide adaptive feature learning. First, we design a Density Generation Branch (DGB) to model object distribution patterns, providing quantifiable priors that guide the network toward dense regions. Second, to address the computational bottleneck of global attention, our Dense Area Focusing Module (DAFM) uses these density maps to identify and focus on dense areas, enabling efficient local-global feature interaction. Finally, to mitigate feature degradation during hierarchical extraction, we introduce a Dual Filter Fusion Module (DFFM). It disentangles multi-scale features into high- and low-frequency components using a discrete cosine transform and then performs density-guided cross-attention to enhance complementarity while suppressing background interference. Extensive experiments on the AI-TOD and DTOD datasets demonstrate that DRMNet surpasses state-of-the-art methods, particularly in complex scenarios with high object density and severe occlusion.

Figures (6)

• Figure 1: Visualization of real-world scenarios with densely arranged tiny objects in remote sensing imagery. (a) Closely arranged cars in a parking lot; (b) Ships navigating along the seashore; (c) Densely packed objects of different categories in a scene. Tiny objects in remote sensing imagery are often densely packed, appearing as scattered clusters.
• Figure 2: Visualization of feature perception of the foreground: The density maps generated by our Density Generation Branch (DGB) exhibit stronger foreground perception capabilities compared to the model's underlying features and can more clearly distinguish object boundary information.
• Figure 3: The architecture of proposed DRMNet. In our network, the Dense Area Focusing Module (DAFM) acquires layers C2 and C3 of the feature pyramid. Based on the density map information corresponding to the object, it obtains the corresponding object feature aggregation regions. This module has relatively low computational cost, yet it achieves sufficient information interaction between focused features and global features. The Dual Filter Fusion Module (DFFM) uses a density-guided cross-enhancement mechanism to fuse these components, effectively suppressing background noise while integrating key details and contextual information.
• Figure 4: AR curve of the AI-TOD test set on different detection backbone networks.
• Figure 5: Visualize learning attention maps using GradCAM. The comparison of results from different methods shows that our approach can focus on object locations better and is less affected by background noise. The density map further enhances the model's ability to focus on object cluster areas.