Answers are grounded in this paper and cite their sources.
Recently, ViTAE-RVSA, the first large-scale Vision Transformer (ViT) tailored for remote sensing, has demonstrated the potential of ViTs by integrating window attention with a convolutional branch. However, as this branch still adopts isotropic square kernels and conventional Feature Pyramid Networks neglect directional priors in multi-scale fusion, the resulting representations remain suboptimal for arbitrarily oriented, high-aspect-ratio, and small objects. To overcome these limitations, we propose Strip-ViTAE, a transformer-based detector that embeds a rotation-equivariant convolution (R2Conv) in the backbone to capture orientation-consistent local features. Moreover, we introduce a Parallel Strip Convolution Module (PSCM) employing horizontal and vertical strip kernels to adaptively enlarge the receptive field and strengthen local feature modeling for elongated and small objects. Finally, we design a Strip Feature Pyramid Network (StripFPN) composed of a Bottom-Up Reflow Module (Bottom-Up RM), a Strip-Based Enhancement Module (SBEM), and a Self-Attention Excitation Module (SAEM) to achieve direction-consistent cross-scale fusion. Experiments on the DOTA-v1.5 dataset show that Strip-ViTAE attains 72.82% mAP, surpassing the state-of-the-art (SOTA) by 1.35% and improving the mAP for high-aspect-ratio and small objects by 3.20%, validating the effectiveness of our method.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: