Improved diffusion model with dedicated pooling and normalized Wasserstein distance for small object detection in UAV images

Abstract

This paper presents an improved diffusion model specifically designed for small object detection in UAV imagery. To mitigate the uneven distribution of noisy bounding boxes across different feature pyramid levels, an inherent drawback of diffusion-based detectors, we design a decoupled region-of-interest pooling method termed Decouple-RoIAlignv2. This method efficiently extracts multi-scale features by applying Alignv2-L to shallow feature maps to enhance fine-grained information and Alignv2-H to deep feature maps to strengthen semantic representation. Meanwhile, the module adopts a strategy of channel-wise concatenation and spatial-wise addition to fuse outputs from different levels, thereby preserving richer detailed information and enhancing complementarity between levels, ultimately improving the joint modeling capability for both small and large objects. To address the defects of classic bounding box loss functions on tiny, sheltered and scale-diverse aerial targets, we introduce a loss term built upon Normalized Wasserstein Distance (NWD). Comprehensive quantitative and visual experiments on two standard aerial benchmark datasets, VisDrone2019 and TinyPerson, validate the superiority of our optimized diffusion detection pipeline for both micro and large-scale objects.