A deep learning-based framework for efficient and accurate 3D real-scene reconstruction

Normally, 3D scenes can be reconstructed with the help of several photo-imaging techniques that are used for 3D geo-visualization and scene analysis. However, the methods involved in the reconstruction of scenes are not accurate and time-consuming. Thus, the given paper proposes a 3D reconstruction framework from multi-view stereo (MVS) images based on deep learning. The proposed multi-view aggregation matching network (MVAMN) comprises pre-processing of the dataset, multi-scale feature extraction using convolutional neural network (CNN), cost volume regularization, and refinement of depth map inference to handle more challenging situations, including mutual occlusions, big-depth variations in oblique photos, and substantial viewpoint changes. The refinement phase involves the computation of probability distributions from the initial depth map as well as the refined depth map. After the refinement is done, the loss function is calculated to quantify the errors between the initial and refined depth maps, thereby optimizing and enhancing the performance of the proposed framework. Lastly, the performance of the model is validated and compared with existing deep learning MVS methods by taking open-source image dataset into consideration. The findings demonstrate that lower memory utilization and higher efficiency make the proposed model highly suitable for aerial images in large-scale and high-resolution 3D surface reconstruction tasks.