A Novel One-Step Small Object Detector for Autonomous Aerial Vehicles

Detecting small objects in aerial images is significantly challenging due to their nonuniform distribution and severe scale variations resulting from changing view angles. Because autonomous aerial vehicles have limited computational power, balancing detection accuracy and efficiency remains a challenging problem. Existing methods, e.g., feature pyramid network (FPN)-based algorithms, concentrate on fusing deep low-resolution features with shallow high-resolution features and primarily rely on simple stacking and channel fusion. However, features of small objects are easily affected by unpredictable noise from the background, leading to high computation cost during feature fusing. This work tackles the issue by designing a novel one-step generative small object detection (SOD) framework. It leverages the self-consistency property provided by a consistency model, which enables the proposed model to convert random Gaussian noise to a single-scale output, thereby enabling the “one-step” inference. We formulate an SOD task as a noise-to-box procedure. We then apply a consistency model to initialize the diffusion process with Gaussian noisy bounding boxes derived from their corresponding ground-truth (GT) annotations. We next introduce a denoising sampling strategy to classify and locate small objects by iteratively refining their Gaussian distributions. We finally comprehensively evaluate our proposed framework on several SOD benchmarks for autonomous aerial vehicles, including DOTA, VisDrone, and AAVDT. Experimental results corroborate that it outperforms the state-of-the-art method (DiffusionDet) by up to 5.1% in terms of AP_{S} (average precision on small objects) on DOTA.