Object 3D Reconstruction Using Photogrammetry and AI

Abstract

This thesis focuses on 3D object reconstruction using photogrammetry combined with artificial intelligence (AI), a rapidly evolving field with uses in the gaming industry, e-commerce, digital archiving, interior designing, and manufacturing. Conventional photogrammetry techniques struggle with noisy data and background clutter, requiring significant manual intervention. To overcome these challenges, we propose an AI-assisted pipeline that improves classical methods to produce accurate and detailed 3D Representations with minimal human effort. Our 3D reconstruction pipeline begins by capturing a 360-degree video of an object. After pre-processing the video, feature extraction and matching are performed using COLMAP, a state-of-the-art structure-from-motion (SfM) and multi-view stereo (MVS) tool. To refine the initial point clouds and address gaps or inconsistencies, we integrate Gaussian Splatting, which helps in creating a smoother and denser representation. Finally, we apply the truncated signed distance function (TSDF) to construct the full 3D geometry and generate a mesh with improved depth consistency and surface quality. Results show that our hybrid approach significantly improves reconstruction accuracy, especially for objects with complex textures and irregular surfaces. Compared to traditional photogrammetry workflows, our method provides cleaner outputs and reduces post-processing time. Finally, we concluded that by combining photogrammetry tools like COLMAP with modern AI-based enhancements such as Gaussian splatting and TSDF function, this project delivers a robust and efficient pipeline for high-quality 3D object reconstruction. Future work involves refining the reconstructed outputs using transformer-based models to further enhance detail, texture consistency, and model realism.