| Περίληψη: |
Highlights: What are the main findings? This work introduces BDNet, a two-path stereo architecture designed for remote sensing disparity estimation. It combines progressive dilation, a decoupled multi-scale reduction module, and strip attention, which cooperate to enhance multi-scale representation and extract directional cues more effectively from urban images. On the US3D dataset, BDNet exceeds all chosen baselines in accuracy with a compact size of 1.43 M parameters and 125.43 G FLOPs. Zero-shot evaluations on Omaha and WHU-Stereo further demonstrate its robust cross-domain capability across various sensors and urban scenes. What are the implications of the main findings? The joint use of progressive dilation and the DMSR module decreases multi-scale feature channels from 320 to 32 while retaining representational capacity. Such compactness enables the network to pull out informative features with low computational cost. The dual-path hourglass structure, combined with attention-guided fusion, enhances disparity estimation in challenging regions such as textureless areas, repeated patterns, and occlusions. This setup achieves a favorable trade-off between accuracy and computational cost. High-resolution remote sensing stereo matching is challenging due to heavy computation and the difficulty of handling textureless areas, repetitive structures, and occlusions. To tackle these issues, we design BDNet (Balancing Dual-path Network)—a stereo matching network that seeks a reasonable trade-off between accuracy and efficiency for remote sensing applications. In the feature extraction stage, BDNet adopts progressive dilation with rates 5, 4, and 3, together with a decoupled multi-scale reduction (DMSR) module, which reduces multi-scale feature channels from 320 to 32. We also introduce a strip attention module to make the network more sensitive to horizontal and vertical structures commonly seen in urban scenes. For cost volume construction, the number of correlation groups is lowered from 40 to 8, in line with the compact 32-channel feature representation. For cost aggregation, a dual-path parallel hourglass architecture is designed, which preserves fine details through a high-resolution path while capturing global context through a low-resolution path. An attention-guided fusion module adaptively integrates features from both paths, improving accuracy in challenging regions such as textureless areas and disparity discontinuities. Experiments on the US3D and WHU-Stereo datasets demonstrate that BDNet achieves the best accuracy among the selected baseline methods on US3D, with D1 errors of 16.05% on Jacksonville and 11.84% on Omaha. It requires only 1.43 M parameters and 125.43 G FLOPs, achieving a favorable balance between accuracy and efficiency. Zero-shot generalization experiments on Omaha and WHU-Stereo further suggest the model's potential for cross-domain adaptation to different satellite sensors and urban scenes. [ABSTRACT FROM AUTHOR] |