Consistent and comprehensive scale aggregation network for drone-view small object detection.

Abstract

Accurate detection in UAV scenarios is challenging due to the small size of objects. This is attributed to two non-negligible factors: (1) small objects lack sufficient visual cues and often suffer information loss during feature extraction; (2) the positional sensitivity of small objects complicates the optimization of predicted bounding boxes. To address these issues, we present a Consistent and Comprehensive Scale Aggregation Network (C²SANet). For high-quality feature representations of small objects, C²SANet develops a novel Multi-Scale Interactive Feature Pyramid Network (MSI-FPN), which introduces two new components based on the top-down propagation path: the Deformable-Based Spatial Calibration (DSC) and Scale Feature Enhancement (SFE) modules. Specifically, DSC leverages pixel-spatial information between adjacent scale features to adjust up-sampled deep features, enhancing the consistency of semantic propagation. SFE first unifies the spatial size of all scale features, then achieves the "Collect-and-Distribute" of full-scale information through the scale interaction block with an encoder-decoder structure, ensuring that the shallow features can be complemented with comprehensive semantic information. Additionally, to improve the localization prediction of small objects, a Coarse-to-Fine Detection Head (CFDH) with geometrical-aware adjustment is devised to refine the quality of predicted boxes iteratively. Extensive experiment results demonstrate the effectiveness and generalizability of C²SANet in improving small object detection performance.