Real World 3D Reconstruction Using Gaussian Splatting

In computer vision, reconstruction is the process of creating a 3D model from image/video frames. 3D Gaussian Splatting (3DGS) is a 3D reconstruction and advanced rendering technique that can create a photorealistic representation of a scene using gaussians (i.e., ellipsoids). This representation allows for the reconstruction of complex real-life environments in extremely high detail. However, 3DGS has its limitations. Our study explored 3DGS and attempted to address its limitations and hypothesize applications. We utilized software such as COLMAP to understand and obtain point clouds, which is also the first step to 3DGS. Then, we used 3DGS with various images and videos we captured around the campus of Clemson University. With these tests, we found out that there is one main limitation to 3DGS: it can only accurately reconstruct static objects. This means that anything that moves, such as tree leaves, cannot be reconstructed with accuracy, but performed with static objects, 3DGS has the potential to reconstruct any complex scene in a very photorealistic manner. One major application to 3DGS is its implementation in interactive simulations (i.e., video games). This displayed another limitation of 3DGS: because 3DGS is a point cloud and no geometric surfaces exist on it, a game cannot be easily created using a gaussian splat. There is ongoing research to address this. Besides these limitations, 3DGS is a cutting edge technique that is much more efficient than its predecessors, and 3DGS will likely revolutionize computer graphics as its shortcomings are addressed.