MambaSOD: Small Object Detection from Drone Perspective in a Mamba-Based Query-Feature Interaction Framework

Drone-perspective small object detection requires processing high-resolution imagery where objects typically occupy fewer than 32×32 pixels, demanding both fine-grained spatial preservation and global context modeling under strict computational constraints. Convolutional neural networks are limited by local receptive fields and cannot effectively model global context, while Transformer-based approaches, despite their global modeling capability, suffer from O(N²) computational complexity that becomes a severe bottleneck when processing high-resolution inputs. This paper proposes MambaSOD, an end-to-end detection framework with an encoder–decoder architecture based on state space models that achieves linear computational complexity throughout the entire pipeline of feature extraction, multi-scale fusion, and query interaction. On the encoder side, MambaSOD builds a linear-complexity multi-scale representation by coupling a Vision Mamba backbone with a P2 high-resolution enhancement module, which recovers fine-grained texture details through dual-path fusion of shallow image features and up-sampled backbone output, together with a BiFPN that performs weighted bidirectional fusion across five scales. On the decoder side, we replace both self-attention and cross-attention with Mamba-driven query interaction: the Mamba-based Query Self-Interaction (MQSI) module enables implicit inter-query communication through bidirectional state propagation at O(N_q) cost, while the Mamba-based Query-Feature Interaction (MQFI) module reformulates query-feature cross-attention as a sequence modeling problem, reducing its complexity from O(N_q × N) to O(N_q + N). Experiments on the VisDrone2019 and UAVDT datasets demonstrate that MambaSOD achieves 23.8% AP, a 52.6% relative improvement over the Vision Mamba baseline, while requiring fewer FLOPs than state-of-the-art detectors such as Cascade R-CNN, ViT, Deformable DETR, and DINO, offering a competitive accuracy-computation trade-off for high-resolution drone-perspective small object detection.