File size: 6,804 Bytes

---
library_name: pytorch
tags:
  - point-cloud
  - semantic-segmentation
  - point-transformer-v3
  - lidar
  - 3d
datasets:
  - RPTU-FGMB/DeKH
license: cc-by-nc-sa-4.0
---

# BIMStruct3D-segmentation

Point-cloud semantic segmentation model (PT-v3m1) from the
[BIMStruct3D](#citation) scan-to-BIM pipeline.

[![License: CC BY-NC-SA 4.0](https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg)](https://creativecommons.org/licenses/by-nc-sa/4.0/) [![arXiv](https://img.shields.io/badge/arXiv-2604.24311-b31b1b.svg)](https://arxiv.org/abs/2604.24311) [![EC3 2026](https://img.shields.io/badge/EC3-2026-blue.svg)](https://ec-3.org/publication/ec32026_197/) [![Dataset](https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-DeKH-ffd21e.svg)](https://huggingface.co/datasets/RPTU-FGMB/DeKH)

This folder is a self-contained package: model weights, config, a minimal copy
of the [Pointcept](https://github.com/Pointcept/Pointcept) codebase (which
this model was trained with), and one script (`segment_scan.py`) that takes a
raw `.las`/`.laz`/`.ply` scan and produces a segmented file of the same
format back out.

## What's in here

```
configs/model_config.py   model architecture + class list
weights/model_best.pth    trained checkpoint
pointcept/                trimmed copy of the Pointcept codebase (see below)
segment_scan.py           chunk -> infer -> merge inference script
pyproject.toml / uv.lock  environment definition
```

## Model

- Backbone: **PT-v3m1** (Point Transformer V3), ~46M parameters.
- Pretrained jointly on Structured3D + ScanNet + S3DIS (Point Prompt
  Training / PPT-v1m1) with a `RandomColorDrop` augmentation (points
  randomly lose their color during training, so the model doesn't over-rely
  on it).
- Fine-tuned on **CV4AEC** training data (indoor/construction point clouds).
- 10 output classes: `clutter, floor, ceiling, wall, column, door, window,
  stairs, railing, lights` (see `configs/model_config.py` -> `data.names`).

The `pointcept/` directory here is deliberately **not** a full checkout of
the Pointcept codebase or of the internal repo this model was trained in —
it's trimmed to only what's needed to build this specific model and run
inference (backbone + generic dataset/transform machinery), so it doesn't
pull in unrelated dataset loaders or the CV4AEC-specific dataset class
(`CV4AECDataset`) at all -- `segment_scan.py` builds tiles as plain
`pointcept.datasets.DefaultDataset` items instead, so you don't need the
CV4AEC dataset or its raw data to run this. If you want the full upstream
codebase instead (e.g. to train further), see the notes at the bottom.

## Setup

Requires a CUDA GPU (tested on an RTX 4090) and
[`uv`](https://docs.astral.sh/uv/):

```bash
uv sync
```

This creates `.venv/` with Python 3.12, PyTorch 2.7 (CUDA 12.6), spconv,
torch-scatter/torch-cluster, and the handful of other packages
`pointcept/` needs -- pinned in `uv.lock` for a reproducible install.
`flash-attn` is deliberately **not** installed (it requires a slow
from-source build); the model falls back to standard attention, which this
config already reflects (`enable_flash=False`).

## Running it

```bash
uv run segment_scan.py --input your_scan.las --output your_scan_segmented.las
```

`--input`/`--output` accept `.las`, `.laz`, or `.ply` (mix and match --
input and output don't need to be the same format).

What it does:
1. Reads the point cloud and estimates per-point normals (the model expects
   color + normals as input features).
2. Splits it into overlapping ~20m tiles (large scans don't fit the model in
   one shot) -- `--tile-size` / `--overlap` to change this.
3. Runs each tile through the model with 10-way test-time augmentation
   (5 scales x flip/no-flip), softmax-averaged.
4. Merges tiles back into one point cloud: for points covered by more than
   one tile (the overlap region), takes a majority vote across tiles.
5. Writes a file with the original coordinates/colors plus a per-point
   `classification` field (0-9, matching the class list above -- for `.ply`
   this is a custom scalar field, readable in CloudCompare/MeshLab) and a
   companion `<output>_labels.npy`.

On a ~6.6M point scan this takes about 8 minutes end-to-end on an RTX 4090
(most of it is the 10x test-time-augmentation passes per tile). Use
`--tile-size` to trade off tile count vs. points-per-tile if you need to fit
tighter GPU memory, or edit `TEST_CFG["aug_transform"]` at the top of
`segment_scan.py` to drop some of the TTA passes for a faster, slightly
less accurate run.

## If you want the full upstream Pointcept instead

This delivery's `pointcept/` is intentionally minimal. If you need the rest
(other model architectures, training scripts, the full dataset zoo), clone
<https://github.com/Pointcept/Pointcept> and drop `configs/model_config.py`
+ `weights/model_best.pth` into it -- the checkpoint loads cleanly against
current upstream (`strict=True`), with one config-level rename to be aware
of: newer Pointcept renamed the backbone's `cls_mode` argument to
`enc_mode` (already reflected in `configs/model_config.py` here).

## License

- **Code** (`segment_scan.py`, `pointcept/`, configs): MIT.
- **Weights** (`weights/model_best.pth`): CC-BY-NC-SA-4.0.

## Citation

If you use this model, please cite the paper it was built for:

```bibtex
@inproceedings{chamseddine2026bimstruct3d,
    author       = {Chamseddine, Mahdi and Kaufmann, Fabian and Schellen, Marius and Glock, Christian and Stricker, Didier and Rambach, Jason},
    booktitle    = {Proceedings of the European Conference on Computing in Construction (EC3)},
    organization = {European Council on Computing in Construction},
    title        = {BIMStruct3D: A Fully Automated Hybrid Learning Scan-to-BIM Pipeline with Integrated Topology Refinement},
    year         = {2026},
    doi          = {10.35490/EC3.2026}
}
```

Please also credit the [CV4AEC Scan-to-BIM Competition 2024](https://cv4aec.github.io/cvpr2024)
([challenge repo](https://github.com/GradientSpaces/cv4aec-challenge), Gradient
Spaces @ Stanford) if you use data or results tied to that benchmark -- this
is the specific challenge whose data the model was fine-tuned on.

## Related dataset

This model was used to produce the semantic segmentation underlying the
BIMStruct3D pipeline's results on the **DeKH (German Hospital) dataset**:
real hospital point clouds (four scenes across three buildings) with
semantic annotations and ground-truth IFC BIM models, released alongside
the paper above.

- Dataset: <https://huggingface.co/datasets/RPTU-FGMB/DeKH>
- License: CC BY-NC-SA 4.0 (non-commercial)

## Acknowledgement

This research was funded by the European Union as part of the projects:
HumanTech (Grant Agreement 101058236) and ShieldBOT (Grant Agreement
101235093).