Question
How should people navigate and understand physical spaces as creative archives?
Context
Over two years my studio and bedroom at CMU accumulated a specific kind of density — things placed, moved, replaced, forgotten. The arrangement of the space was an ongoing record of how institutional life reshapes daily habits. At the end, before I left, I scanned it.
Experiment & technical approach
Capture used Scaniverse on an iPhone (LiDAR + visual SLAM), with 3D Gaussian Splatting trained from the resulting point cloud. Alternative capture processes also used include capturing raw images of the studio and bedroom with a Fujifilm X-T5 and training a 3D Gaussian Splatting model, then converting to a splat representation. The harder problem was making the result navigable and legible. I wanted each object to be queryable — not labeled by hand, but surfaced through the geometry of the space itself.
The segmentation pipeline weighs surface normals at 2.5× position features to identify object boundaries in a densely-packed scene, then applies HDBSCAN clustering, CLIP for non-taxonomic semantic queries, and SAM for 2D instance masks. Visibility scoring prunes floater Gaussians in free space. Hand segmentation (drawing bounding boxes) was also employed to refine the object boundaries. The result is a 3D archive of the space, where each object can be queried and explored in a browser.
The interface avoids gallery conventions: it's a command-line-style filesystem with typewriter reveals and first-person navigation. Objects appear as evidence, not displays. Conceptually the work sits with Tracey Emin's My Bed and Gala Porras-Kim's archival practice — the idea that a room, documented thoroughly enough, becomes a primary source.
Findings
The capture itself was easier than making a dense, personal room legible. Fully automatic clustering confused overlapping objects and produced floating fragments, while hand-authored labels erased the ambiguity that made the archive interesting. Combining geometric cues, semantic models, and selective manual correction produced a better division of labor.
Implications
Spatial archives need interfaces for inquiry, not only photorealistic walkthroughs. Creative tools for 3D collections could let people search by association, move between object and room scale, and preserve uncertainty rather than forcing every object into a fixed museum taxonomy.
documentation
