91% SEGMENTATION · −23% DEPTH ERROR · UNDER 90s / SCENE
One photograph in. A room you can walk through, out.
ATRIUM takes a single equirectangular (2:1) panorama, reconstructs a metric, textured 3D mesh of the room on a free Kaggle T4 GPU, and opens it as a first-person browser walkthrough — WASD and mouse on desktop, touch on mobile. No point clouds to clean, no photogrammetry rig, no paid compute anywhere in the loop.
INPUT
1
panorama, nothing else
SEGMENTATION
91%
indoor + outdoor scenes
DEPTH ERROR
−23%
vs off-the-shelf baselines
GPU COST
$0
Kaggle T4, orchestrated by API
A / THE PIPELINE
Local orchestration, borrowed GPU.
B / ENGINEERING NOTES
Learned the hard way.
- T4, never P100. Kaggle's preinstalled PyTorch dropped the P100's sm_60 compute capability; the T4 (sm_75) runs the stack as-is. One line of kernel metadata is the difference between a working pipeline and a silent crash.
- One forward pass, no seams. UniK3D is run with its native Spherical camera over the entire equirectangular image — metric 3D points out, no cube-face stitching artifacts. Installed with --no-deps so pip can't disturb Kaggle's torch build.
- The mesh is the depth map. A vertex grid is lifted straight from predicted depth: the seam column is duplicated for a clean UV wrap, stretch triangles are culled by relative depth ratio, and the nadir is flattened onto the detected floor plane.
- metallicFactor = 0. glTF defaults metalness to 1.0, which renders pitch black in an unlit viewer. The walkthrough uses unlit textured materials with tone mapping off — the panorama's own light is the lighting.