3D GAUSSIAN SPLAT (3DGS) of Jack Lenor Larsen's room
Captured in 2024 at LongHouse Reserve.
Emil Polyak, Kathi Martin
The project integrates multiple computational photogrammetry and point-based rendering techniques to explore alternative methods of archival 3D capture. It demonstrates how point-based rendering, particularly 3DGS, can complement traditional photogrammetry for preserving complex interior spaces.
​
Scene Description and Challenges
The room contains a diverse range of objects, materials, and textures, which poses challenges for accurate photogrammetric reconstruction. Lighting conditions included natural light from three windows (softened by outdoor vegetation) and two softbox artificial light sources for fill lighting. These mixed lighting sources created variable illumination and soft shadows. The complexity of the scene required careful camera movement during capture and extensive data processing to ensure all details were recorded without significant artifacts.
Image Acquisition
A Sony A7R IV camera with a 24 mm lens was used to capture a 4K video of the space. Video frames were extracted at 1 FPS, yielding 1,646 frames using Shutter Encoder (Shutter Encoder, 2024). This provided a dense set of images while maintaining manageable computational overhead for processing. This mosaic of frames illustrates the extensive coverage of the room achieved by moving the camera throughout the space. The consistent overlap and varied angles in the frames helped ensure that all objects and surfaces in the room could be reconstructed during 3D processing.
 |  |  |
|---|---|---|
 |
Sparse point cloud of Jack Lenor Larsen’s room from COLMAP’s reconstruction. The red frusta represent the camera positions and orientations for the 1,646 video frames. Approximately 783,000 points were triangulated, capturing the major structures and objects in the room. The sparse model’s mean reprojection error was around 0.95 pixels, indicating a good initial alignment between the points and the input images.
Processing Workflow in COLMAP
The 1,646 extracted frames were imported into COLMAP for sparse reconstruction. Feature detection and matching were performed, followed by bundle adjustment to solve for camera poses. The following key settings were used in COLMAP for SfM:
Feature Extraction: SIMPLE_RADIAL camera model with a maximum of 8,192 features per image.
Feature Matching: Sequential matching with image overlap of 10, matching ratio test at 0.75, and guided matching enabled for consistency.
Sparse Reconstruction: Bundle adjustment refining focal length, with a minimum of 20 inliers per image to accept camera poses.
​
After processing, COLMAP produced a sparse point cloud and estimated camera trajectories for all frames. The reconstruction statistics indicate 1,646 registered images, about 783,718 sparse points, and a mean reprojection error of 0.95 pixels. This level of accuracy is a strong foundation for the subsequent dense modeling steps.
Gaussian splat rendering of the room after processing with PostShot. The 3DGS model contains approximately 2.75 million point-like “splats” that encode color and brightness. A spherical harmonics level of 3 was used for each splat to capture lighting effects, resulting in smooth yet detailed shading across surfaces. The PostShot optimization ran for 183,000 iterations (kSteps), reaching an SSIM of 0.811, which reflects high fidelity to the original imagery.
Converting COLMAP Output to 3DGS with PostShot
The sparse COLMAP output (camera poses and sparse points) was then converted into a dense 3D Gaussian Splatting model using PostShot v0.5.190 (PostShot, 2024), an experimental Gaussian splatting tool. PostShot uses the camera poses to optimize a set of 3D Gaussian primitives (splats) that reproduce the input images. The COLMAP data (exported as binary .bin files for cameras and points) was fed into PostShot to initialize this process.
After running for about 183 kSteps of optimization, PostShot yielded a dense splat-based reconstruction with roughly 2.75 million Gaussian splats. The model utilized spherical harmonics of degree 3 for each splat, enabling smooth shading while retaining detail from the original photographs. The final training quality reached an SSIM (Structural Similarity Index) of 0.811, indicating the reconstructed splat model is structurally and visually similar to the source images.
Point cloud after applying outlier removal, noise filtering, and leveling in CloudCompare. The filtered point cloud is much cleaner: most of the spurious points have been removed, and the overall structure of the room is clearer. Walls, furniture, and objects are better defined, and the scene is properly aligned with the horizontal plane, which will benefit visualization and measurement accuracy.
Point Cloud Filtering in CloudCompare
The raw splat model was converted to a standard point cloud format (PLY) using the 3DGS Converter tool (Fugazzi, 2024) for further cleaning. This dense point cloud was imported into CloudCompare (CloudCompare, 2024) to refine the data by removing noise and outliers. Even with careful capture and 3DGS processing, some spurious points can appear (for example, floating points in empty areas or around object edges). In the unfiltered model, a number of outlier points and noise artifacts are visible around the room, especially near high-contrast edges and in empty space. These errors are common in dense reconstructions and need to be removed to improve clarity. To clean the point cloud, a Statistical Outlier Removal (SOR) filter was applied to eliminate stray points that deviate from their neighbors. This was followed by additional noise filtering to remove isolated clusters of points that do not belong to any surfaces (such floating artifacts often appear as speckles in mid-air). Finally, the point cloud was leveled to align the floor and ceilings to the horizontal plane, ensuring the model is correctly oriented in world coordinates.
Final Interactive Visualization in Supersplat
The final interactive visualization of the room was achieved using Supersplat, allowing real-time exploration of the space. The Gaussian splat model is rendered directly in a web browser with correct colors and lighting. Viewers can orbit and zoom to examine details of the room, demonstrating how a complex real-world interior can be shared online for educational and preservation purposes.
While the results effectively capture the intricate details of the space, some limitations remain, such as residual noise and floating splats, particularly in areas that were challenging to capture from multiple angles. For example, due to the open ceiling, some structures were not visible from enough perspectives, leading to incomplete or less stable reconstructions. Despite these challenges, the continuous and rapid advancement of Gaussian Splatting technology is highly promising, offering new possibilities for more complete and accurate digital preservation.
​
Conclusion
This project demonstrates the viability of 3D Gaussian Splatting for digital preservation, offering an efficient and visually accurate alternative to traditional photogrammetry. By capturing a dense set of images and leveraging tools like COLMAP, PostShot, and Supersplat, we reconstructed Jack Lenor Larsen’s room with remarkable detail.
The point-based approach bypasses the need for explicit mesh modeling, directly rendering from Gaussian splats, which simplifies the workflow for complex, irregular scenes.
Overall, 3DGS has proven to be a promising method for archival 3D capture, particularly for preserving the visual richness of artistically significant spaces such as this one. With ongoing advancements in point-based rendering techniques, the potential for high-fidelity, immersive digital preservation continues to expand.
References
CloudCompare. (2024). Open-source 3D point cloud and mesh processing software. Retrieved from https://www.cloudcompare.org/
Fugazzi, F. (2024). 3DGS Converter [Computer software]. Retrieved from https://github.com/francescofugazzi/3dgsconverter
PostShot. (2024). Pre-release beta v0.5.190 [Computer software].
Schönberger, J. L., & Frahm, J. M. (2016). Structure-from-Motion Revisited. Proceedings of CVPR, 4104–4113.
Shutter Encoder. (2024). Free video and image processing software. Retrieved from https://www.shutterencoder.com/
Supersplat. (2024). Interactive Gaussian Splatting renderer for web visualization [Computer software].