Multiscale Dynamic Receptive Field for High-Resolution Satellite Stereo Matching

Authors:
Publisher website
Google Scholar
Abstract
High-resolution satellite stereo matching is fundamental for 3D mapping. While current deep learning models achieve impressive macroscopic geometric accuracy, they inherently rely on convolutions with static receptive fields. This rigidity struggles with extreme scale variations across complex landscapes, often causing pervasive localized matching errors, such as edge-fattening artifacts at object boundaries. Furthermore, applying standard dynamic convolutions to address this within massive 4D cost volumes incurs prohibitive memory overhead. To address this trade-off, we propose a parameter-efficient, scale-adaptive cost aggregation strategy. We introduce a Multi-scale Dynamic Receptive Field (MDRF) mechanism that decouples the learned base operator from the dilation-dependent sampling support. By reusing a single shared kernel across multiple dilations via a content-aware attention gate, the network adaptively weights multi-scale context responses. This plug-and-play module is parameter-efficient, with memory and throughput explicitly reported in the evaluated configurations. Evaluated on the WHU-Stereo and US3D benchmarks, our approach consistently lowers the outlier rate, achieving absolute D1-error reductions of 0.43 and 0.49 percentage points, respectively, compared with the baseline. Error-map analysis shows that MDRF reduces localized boundary artifacts, supporting a measured accuracy-efficiency trade-off for practical remote sensing applications.

This publication has 25 references indexed in Scilit: