Multi-view stereo-based 3D modeling of waste pits in waste-to-energy plants with spatial-context feature fusion

Three-dimensional (3D) modeling of waste pits can support waste pile monitoring and crane-assisted feeding in waste-to-energy (WtE) plants. RGB image-based multi-view stereo (MVS) offers a flexible perception route, but its performance remains limited by non-uniform illumination, weakly textured waste surfaces, local specular highlights, occlusion, and changes in pile morphology. This study aims to improve the reconstruction continuity and relative accuracy of RGB-MVS under these waste-pit-oriented visual conditions. To this end, SCFM-MVSNet is proposed by introducing a Spatial-Context Fusion Module (SCFM) into multi-scale feature extraction and a spatial-confidence-guided Gated Recurrent Unit (GRU) strategy into iterative depth refinement. The SCFM enhances spatial-context feature representation, while the confidence-guided GRU reduces the influence of unreliable matching features during depth updating. Experiments on public benchmarks show that SCFM-MVSNet achieves an Overall error of 0.290 on DTU and competitive performance on Tanks and Temples. Tests on self-constructed simulated and real waste pit datasets further show more continuous waste-pile surfaces and lower relative reconstruction errors than the compared methods. These results suggest that spatial-context feature fusion and confidence-aware refinement are beneficial for reducing local holes and unstable matching responses. The main contribution and novelty of this work lie in a waste-pit-oriented RGB-MVS framework that adapts RGB-based MVS to challenging waste pit 3D modeling through spatial-context-aware feature fusion, confidence-aware depth refinement, and paired simulated–real waste pit datasets. Further validation with independent geometric references is still required before autonomous crane-control deployment.