---
license: cc-by-nc-4.0
tags:
- 3d
- point-cloud
- indoor-scene-understanding
- scene-reconstruction
library_name: roomform
---

# roomform patch-graph ConvFormer

Boundary models for [roomform](https://github.com/johnathanchiu/roomform):
point cloud indoor scans in, room boundaries (walls, floors, ceilings —
inferred through occlusion) out, as a patch graph over an 8 cm voxel
lattice: 3 surface node classes + 13 forward-edge connectivity
channels. See the repo's
[model docs](https://github.com/johnathanchiu/roomform/blob/main/roomform/model/README.md)
for the architecture.

## Checkpoints

| file | params | input | heads | notes |
|---|---|---|---|---|
| `patch-graph-joint-rgb-55m-offset-head-r2.pt` | 55M | RGB, 11ch | offsets + door/window openings | **default** |
| `patch-graph-400-offset-isolated-r1.pt` | 27M | grayscale, 9ch | offsets | lightweight |

Format: `{"model": state_dict, "epoch": int, "config": {...}}` — the
config dict loads directly via `roomform.model.config.ModelConfig`.

## Usage

Drop a checkpoint into `checkpoints/` in the repo (the pipeline
downloads the default automatically) and run:

    uv run python -m roomform.pipe.e2e SCAN.ply

## License

CC BY-NC 4.0 — free for research with attribution; commercial use
requires a separate license (see the repo's LICENSE-WEIGHTS and
CITATION.cff: Johnathan Chiu, Matthew Zhou, Preston Bourne).
Trained entirely on synthetic data generated by the roomform internal
data pipeline; no third-party dataset terms attach to the weights.