DeCoF: Tiny Object Detection in Aerial Remote Sensing Imagery via Detail-Retentive Downsampling and Conflict-Aware Fusion

Abstract

Aerial object detection has witnessed remarkable progress, yet detecting tiny objects in remote sensing imagery remains highly challenging due to their extremely limited spatial footprints. Hierarchical feature processing pipelines tend to exacerbate this problem, as their standard spatial downsampling operators discard fragile structural cues, while cross-scale fusion often overlooks the representational inconsistency between shallow details and deep contextual priors. To address these bottlenecks, we propose DeCoF for tiny object detection in aerial imagery via two complementary modules. First, we introduce the detail-retentive downsampling (DRD) module, which performs spatial downsampling through polyphase decomposition followed by detail re-encoding. Equipped with learnable structure-preserving shrinkage and a residual-guided update mechanism, DRD suppresses background clutter while alleviating the distortion of fragile object cues. Second, we present the conflict-aware fusion (CAF) module, which explicitly measures the representational inconsistency between shallow details and deep contextual priors to guide cross-scale fusion. Specifically, CAF quantifies such discrepancies via Kullback-Leibler (KL) divergence, dynamically highlighting ambiguous regions and selectively enhancing complementary contextual priors for more effective multiscale feature aggregation. Extensive experiments on the VisDrone2019 and AI-TOD benchmarks demonstrate the effectiveness and superiority of the proposed DeCoF. Specifically, on the VisDrone2019 benchmark, integrating DeCoF yields absolute improvements of 1.7 points in mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> and 1.6 points in mAP<inline-formula><tex-math notation="LaTeX">$_{\text{S}}$</tex-math></inline-formula> over the pre-trained YOLOv9 baseline, and 2.0 points in mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> and 1.6 points in mAP<inline-formula><tex-math notation="LaTeX">$_{\text{S}}$</tex-math></inline-formula> over the RT-DETR baseline. Furthermore, on the AI-TOD dataset, the RT-DETR-based DeCoF achieves 51.1% in mAP<inline-formula><tex-math notation="LaTeX">$_{50}$</tex-math></inline-formula> and 22.1% in mAP<inline-formula><tex-math notation="LaTeX">$_{\text{S}}$</tex-math></inline-formula>. These consistent gains across both CNN-based and transformer-based detectors further verify the paradigm-agnostic nature of DeCoF.