Enhancing Archeological Building Reconstruction Through Image-Based 3D Modeling and LoFTR

Archaeological research often relies on the meticulous reconstruction of historical structures and artifacts. However, this process is frequently hindered by the absence of detailed 3D data and the limited availability of well-preserved sites. In response to these challenges, this thesis tries to contribute in rendering the 3D structure of archaeological buildings from 2D unstructured image. In case of damage to the original buildings, its 3D model can be helpful for replication. The images from the internet are taken to train the model using Structure From Motion (SfM) based algorithms. A model named Detector-Free Local Feature Matching with Transformers (LoFTR) is utilized for local image feature matching. Ultimately, the feature points are adjusted to generate a 3D representation of the building. The key stages of structure from motion encompass feature detection and matching, outlier removal, camera pose estimation, point cloud creation, and 3D mesh generation. This reformulation method can reduce the expense and intricacy associated with traditional approaches.