Context-aware depth estimation for improved 3D reconstruction of homogeneous indoor environments

In the architectural, engineering, and construction (AEC) industry, 3D reconstruction is crucial for applications such as construction management, indoor navigation, and energy performance analysis. However, indoor environments, characterized by textureless surfaces and varying lighting conditions, pose significant challenges that conventional reconstruction methods struggle to address effectively. To tackle these issues, this paper proposes IndoCAFE-Net, a deep learning-based Multi-view Stereo (MVS) framework designed to enhance the accuracy and completeness of indoor 3D reconstructions. Trained on the indoor-specific IndoReal-MVS dataset, which captures intricate indoor phenomena such as dynamic lighting and homogeneous areas, IndoCAFE-Net integrates a Context-Aware Feature Enhancement (CAFE) block and a dual-loss optimization strategy. It achieves an accuracy of 4.70 mm, completeness of 5.20 mm, and a Relative Improvement in Valid Points (RIVP) score of 265.16 % over existing models. These results highlight IndoCAFE-Net's potential to advance indoor 3D reconstruction, enabling robust solutions for facility management and asset optimization in the AEC industry.