Research on 3D Reconstruction Technology of Indoor Buildings Based on Depth Prediction

Abstract

To address the limitations of traditional 3D indoor scene reconstruction methods in resource-constrained environments, this paper proposes a depth prediction-based 3D reconstruction method for indoor buildings. The method first employs a pre-trained image encoder to extract multi-scale features from input images, which are then combined with metadata containing ray direction, depth information, and relative pose distance to construct a feature volume. This volume is fed into a 2D convolutional neural network, while a multi-scale depth prediction strategy is adopted to progressively refine depth estimation, generating high-quality depth predictions for more detailed 3D reconstruction. Experimental results demonstrate that the proposed method significantly outperforms traditional depth estimation approaches on the public dataset ScanNet, achieving a 21% improvement under the threshold accuracy metric δ < 1.05. In 3D reconstruction tasks, the method achieves near state-of-the-art performance (F-Score = 0.658) while enabling online real-time reconstruction with low memory consumption, exhibiting a per-frame latency of only 72ms.