RGB-D-Based 3D Scale Recovery for Residential interior Reconstruction

Three-dimensional (3D) visual generation technologies, represented by Neural Radiance Fields (NeRF), have been widely applied in various fields in recent years. Prior to NeRF reconstruction based on deep learning, traditional algorithms are typically used to estimate camera poses. However, Multi-View Stereo (MVS) algorithms solely relying on RGB photographs can only determine the relative 3D structure of targets and cannot accurately estimate their actual 3D scale. Depth cameras, as imaging devices capable of capturing both target pixel information (RGB) and depth information (D), have become popular in civilian applications. MVS technology based on RGB-D images can provide camera poses with real-scale information. However, due to practical factors such as lighting and surface reflection of materials, depth values measured by civilian depth cameras often contain certain errors. Therefore, how to effectively eliminate depth errors in RGB-D images to accurately recover the actual 3D scale of targets is an urgent issue to be addressed in practical applications. To address this issue, this study proposes two improved algorithms for RGB-D multi-view scale recovery based on previous work: The first improved algorithm (Adaptive-3D-BA) directly obtains camera pose estimates with real-scale information by adaptively adjusting depth values and incorporating them into the traditional Bundle Adjustment (BA) algorithm. The second improved algorithm (BA-SE) adds a Scale Estimation (SE) postprocessing module to the traditional BA based on RGB images, utilizing the least squares method to recover the actual scale information of the camera. Comparative analysis on the TUM dataset verifies the effectiveness of the above methods in improving measurement accuracy. The results indicate that the proposed algorithms can provide accurate scale information for indoor 3D reconstruction of old residential buildings, holding broad application prospects.