School of Engineering Department of Computer Science and Engineering 96 AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: CHAN Cheng Leong / COMP Course: UROP 1100, Spring This report summarizes my semester’s research on 3D reconstruction using Fast3R, including both literature review and practical implementation. In the first phase, I reviewed papers such as COLMAP-Free 3D Gaussian Splatting and Fast3R: Towards 3D Reconstruction 1000+ Images in One Froward Pass to understand the underlying techniques of 3D reconstruction techniques. The second phase included implementing Fast3R on a virtual machine to generate and merge point clouds from multi-view images, reconstructing the 3D scene. The results show that Fast3R may prodcues noisy 3D models when fusing all point clouds together, but the reconstruction can be improved by reducing the number of views and removing points with lower confidence, the latter being good for single object reconstruction. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: CHEN Jhao Song / DSCT Course: UROP 1100, Fall This is a comparison across 19 different approaches in missing data imputation. These approaches are examined by 15 datasets with the 3 simulated randomness mechanisms. The result is then analyzed in different aspects aiming to find an overall better strategy. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: CHENG Zi Hao / CPEG Course: UROP 1100, Spring This report presents our preliminary investigation into reconstructing large-scale indoor scenes from sparse, crowdsourced images using state-of-the-art deep learning techniques. Under the supervision of PhD student Sizhe Song, we explore neural radiance field methods (NeRF) implemented via the Nerfstudio framework, as well as explicit point-based methods such as 3D Gaussian Splatting (3DGS) and its variants. Specifically, we test the original 3DGS approach, hierarchical-3D-Gaussians, and recent large-scene adaptations like VastGaussian and CityGaussian. We also consider recent techniques that relax the requirement of known camera poses, such as COLMAP-Free 3DGS, and brainstorm ideas for reconstructing scenes when only partial pose parameters (e.g. 2D camera position) are available. Our experiments use a sparse image dataset collected in the atrium of our university’s campus building. We evaluate reconstruction quality using view synthesis metrics and by visualizing the reconstructed Gaussian point clouds. Preliminary results show that explicit Gaussian-based methods often outperform NeRF. Ongoing work includes applying our dataset to new models like InstantSplat or MVSGaussian. This report outlines background, methodology, experimental results, and future directions.
RkJQdWJsaXNoZXIy NDk5Njg=