FDQ-Det: Frequency-Aware Spatial Modeling and Density-Guided Query for Tiny Object Detection

Tiny object detection (TOD) is hindered by limited feature representation, pervasive background clutter, and extreme density variations. Although Transformer-based detectors benefit from global context, their object queries often suffer from instability and redundancy in dense scenes. This is primarily due to susceptibility to high-frequency interference and a lack of explicit spatial guidance. Motivated by spectral and error attribution analyses, we propose FDQ-Det, a framework designed to stabilize query behavior through frequency-aware spatial modeling and density-guided construction. Specifically, a Count Density Spatial Prior (CDSP) module utilizes continuous density estimation to provide structured spatial guidance for query initialization. Complementarily, a High-Frequency Suppression (HFS) mechanism is developed to attenuate detrimental spectral components, thereby refining localization features. To ensure robustness in crowded scenarios, we further introduce an adversarial query-level regularization to prevent over-confident coupling between density cues and query representations. Extensive experiments on AI-TODv2 and VisDrone benchmarks demonstrate the effectiveness of FDQ-Det. On the challenging AI-TODv2 dataset, our method outperforms strong baselines by 1.9% mAP and 2.5% AP \({}ᵥₜ\) , confirming the efficacy of integrating density-informed priors with frequency-aware feature modeling.