Indigenous Architectural Reconstructions in Palestine

Many historical buildings have been lost or altered over time, leaving only fragmented visual records in photographs or paintings. To digitally preserve architectural heritage, we need methods that can reconstruct buildings from limited visual inputs. While modern computer vision has made progress in generating multiple views and 3D reconstructions from single images, applying these techniques to historical imagery presents unique challenges such as image quality, perspective limitations, and missing data. Last semester, a Spark team built an end-to-end proof-of-concept single-image 3D reconstruction pipeline. The system generates multiple novel views of a building from one photo, estimates depth for each view, converts depth maps into point clouds, and merges them into an exportable 3D model. This semester, the team will build on that foundation to improve quality, reliability, and usability of the reconstruction outputs. The focus for this semester will be on developing and testing the image generation models, while the long-term goal is to use these generated views as the basis for forming 3D building models. If more images are needed, we also have a ready-to-use Google API script that can collect building images from multiple angles based on geographic coordinates. What is the possibility of expanding the scope to the neighborhood, multiple buildings, street view, and diverse urban forms? While capturing its evolution through time? Based on images such as maps and aerial images, By the end of the semester, the team will aim to: • Improve control and consistency of AI-generated novel views. • Reduce geometry artifacts and improve mesh / point cloud cleanliness. • Increase robustness to real-world images (clutter, multiple buildings, imperfect framing). • Move toward a more automated, seamless workflow from input image to exportable 3D asset.