Abstract:We propose C2B-DeformYOLO, an end-to-end real-time UAV road damage detector built on YOLOv10, with targeted architectural improvements for two core issues: cross-scale discontinuous representations of elongated cracks and accurate localization of nonconvex pothole boundaries. First, we introduce a Continuous Conditioned Gather Distribute neck, termed C2GD Neck, which performs global information gathering over multi-level features and generates continuous conditioning weights to adaptively allocate and fuse features in a bidirectional manner. This design strengthens coherent responses for thin crack patterns and enhances long-range dependency modeling across scales. Second, we present Budgeted Offset Sparse Convolution, called BOS Conv, which conducts sparse offset sampling and reconstruction on high-response regions under a constrained computation budget, improving geometric alignment and boundary deformation modeling while preserving efficient inference. Experiments on a UAV road damage dataset and the public RDD2022 benchmark demonstrate consistent gains in detection accuracy and cross-domain generalization over the baseline, without sacrificing real-time performance.