// NSF REU HUMANS MOVE

I worked on a project to compare different methods of 3D reconstruction, focusing on how accurately they could recreate a small, detailed object like the Key Marco Cat. I used photogrammetry and Gaussian Splatting, 3D-printing the original model to use as a baseline for comparison. While I haven’t yet been able to fully test Neural Radiance Fields (NeRF) due to some technical hurdles, I’m actively working on implementing it and hope to incorporate it into this research soon. The main goal of this project is to balance accuracy, efficiency, and practicality across these methods.

I spent a lot of time reading about different 3D reconstruction techniques. Photogrammetry is a go-to for many projects and generally produces reliable results, though it can struggle with fine details and lighting inconsistencies. Gaussian Splatting is a newer, faster method that’s gaining attention for its efficiency and ability to handle complex scenes. NeRF seems particularly promising for capturing intricate details and nuanced lighting, but it’s also highly resource-intensive and technically challenging to implement. While most research focuses on large-scale objects, applying these methods to small, intricate artifacts felt like something worth exploring.

Small, detailed objects present unique challenges for 3D reconstruction, especially when it comes to capturing fine details accurately. This project aims to answer two questions: How well do photogrammetry and Gaussian Splatting recreate a small object compared to its original 3D-printed version? And how might NeRF, once implemented, change the equation in terms of detail and computational cost?

I took three sets of photographs of the Key Marco Cat, varying location and lighting. I preprocessed the images by removing backgrounds for a cleaner input. I then processed the data using RealityCapture (photogrammetry) and Gaussian Splatting to create reconstructed models. A 3D-printed original model to acted as a tangible baseline for accuracy comparisons. The evaluation of the reconstructions was done using SSIM (Structural Similarity Index) to measure how closely each method matched the original. NeRF is still in progress, and I plan to integrate it into the workflow for future comparisons.

Photogrammetry delivered highly accurate reconstructions, closely matching the 3D-printed baseline in terms of detail and structure, though it required significant time and computational resources. Gaussian Splatting, while not as precise, was much faster and less demanding, making it a strong alternative when speed and resource efficiency are priorities. It also created some rather funky models, which I believe might just be user error. Once NeRF is fully implemented, I anticipate it could offer even more detailed results, though likely at a higher computational cost.

This project highlights the trade-offs between accuracy and efficiency in 3D reconstruction methods. Photogrammetry set the standard for detail, but its processing demands were substantial. Gaussian Splatting offered a practical alternative, providing good results in less time. While NeRF hasn’t yet been integrated into this project, its potential to handle complex lighting and detail makes it an exciting avenue to explore further. Using the 3D-printed baseline has been key to objectively evaluating these methods, and I’m looking forward to refining and expanding this research as I continue troubleshooting NeRF.

Even without NeRF fully operational yet, this project has already provided valuable insights into how different 3D reconstruction methods compare, particularly for small-scale artifacts. Photogrammetry remains the gold standard for accuracy, while Gaussian Splatting shows promise for faster, more resource-friendly workflows. I’m eager to see how NeRF might fit into this equation once it’s fully implemented and tested.