CDA-MVSNet: Enhancing asteroid 3D reconstruction with channel attention and dynamic aggregation

Due to the lack of prior knowledge and clear surface textures, it is difficult to accurately reconstruct the three-dimensional model of an asteroid, and there are few scholars engaged in research in this area. In this paper, we propose an Enhancing Asteroid 3D Reconstruction Framework with Channel Attention and Dynamic Aggregation (CDA-MVSNet) to reconstruct the three-dimensional model of an asteroid with monotonic or indistinct textures. This framework enhances feature expressiveness via the channel attention Adaptive Selective Channel Focus (ASCF) module and employs a Depth-Aware Adaptive Loss Function (DALOSS) that dynamically adjusts during iterations to guide the depth estimation finely. The cost aggregation Dynamic Weighting Synthesis Network module (DWSN) we propose further refines the cost aggregation progressively, substantially improving reconstruction precision and robustness. Visualization evaluation conducted on Sample Consensus Initial Alignment Iterative Closest Point (SAC-IA-ICP), Chamfer Distance metrics, and the DTU dataset visually substantiated the superiority of our approach. CDA-MVSNet achieves significant advancements in asteroid 3D reconstruction accuracy and computational efficiency compared to existing methods. Our framework achieves new SOTA performance on reconstruction completeness, a critical metric for our target application. Our method demonstrates strong overall performance on asteroid datasets, validating its effectiveness for this challenging domain.