The P³ Dataset: Pixels, Points and Polygons
for Multimodal Building Vectorization

Name: PixelsPointsPolygons
Creator: Raphael Sulzer
License: https://choosealicense.com/licenses/cc-by-4.0/

Raphael Sulzer^1,2 Liuyun Duan¹ Nicolas Girard¹ Florent Lafarge²

¹LuxCarta Technology
²Centre Inria d'Université Côte d'Azur

Figure 1: A view of our dataset of Zurich, Switzerland

Abstract

We present the P³ dataset, a large-scale multimodal benchmark for building vectorization, constructed from aerial LiDAR point clouds, high-resolution aerial imagery, and vectorized 2D building outlines, collected across three continents. The dataset contains over 10 billion LiDAR points with decimeter-level accuracy and RGB images at a ground sampling distance of 25 cm. While many existing datasets primarily focus on the image modality, P³ offers a complementary perspective by also incorporating dense 3D information. We demonstrate that LiDAR point clouds serve as a robust modality for predicting building polygons, both in hybrid and end-to-end learning frameworks. Moreover, fusing aerial LiDAR and imagery further improves accuracy and geometric quality of predicted polygons. The P³ dataset is publicly available, along with code and pretrained weights of three state-of-the-art models for building polygon prediction at https://github.com/raphaelsulzer/PixelsPointsPolygons.

Highlights

A global, multimodal dataset of aerial images, aerial LiDAR point clouds and building outline polygons, available at huggingface.co/datasets/rsi/PixelsPointsPolygons
A library for training and evaluating state-of-the-art deep learning methods on the dataset, available at github.com/raphaelsulzer/PixelsPointsPolygons
Pretrained model weights, available at huggingface.co/rsi/PixelsPointsPolygons
A paper with an extensive experimental validation, available at arxiv.org/abs/2505.15379

Dataset

Overview

Download

The recommended and fastest way to download the dataset is to run

pip install huggingface_hub
python scripts/download_dataset.py --dataset-root $DATA_ROOT

Optionally you can also download the dataset by running

git lfs install
git clone https://huggingface.co/datasets/rsi/PixelsPointsPolygons $DATA_ROOT

Both options will download the full dataset, including aerial images (as .tif), aerial lidar point clouds (as .copc.laz) and building polygon annotaions (as MS-COCO .json) into $DATA_ROOT . The size of the dataset is around 163GB.

Structure

📁 Click to expand dataset folder structure

PixelsPointsPolygons/data/224
├── annotations
│   ├── annotations_all_test.json
│   ├── annotations_all_train.json
│   └── annotations_all_val.json
│       ... (24 files total)
├── images
│   ├── train
│   │   ├── CH
│   │   │   ├── 0
│   │   │   │   ├── image0_CH_train.tif
│   │   │   │   ├── image1000_CH_train.tif
│   │   │   │   └── image1001_CH_train.tif
│   │   │   │       ... (5000 files total)
│   │   │   ├── 5000
│   │   │   │   ├── image5000_CH_train.tif
│   │   │   │   ├── image5001_CH_train.tif
│   │   │   │   └── image5002_CH_train.tif
│   │   │   │       ... (5000 files total)
│   │   │   └── 10000
│   │   │       ├── image10000_CH_train.tif
│   │   │       ├── image10001_CH_train.tif
│   │   │       └── image10002_CH_train.tif
│   │   │           ... (5000 files total)
│   │   │       ... (11 dirs total)
│   │   ├── NY
│   │   │   ├── 0
│   │   │   │   ├── image0_NY_train.tif
│   │   │   │   ├── image1000_NY_train.tif
│   │   │   │   └── image1001_NY_train.tif
│   │   │   │       ... (5000 files total)
│   │   │   ├── 5000
│   │   │   │   ├── image5000_NY_train.tif
│   │   │   │   ├── image5001_NY_train.tif
│   │   │   │   └── image5002_NY_train.tif
│   │   │   │       ... (5000 files total)
│   │   │   └── 10000
│   │   │       ├── image10000_NY_train.tif
│   │   │       ├── image10001_NY_train.tif
│   │   │       └── image10002_NY_train.tif
│   │   │           ... (5000 files total)
│   │   │       ... (11 dirs total)
│   │   └── NZ
│   │       ├── 0
│   │       │   ├── image0_NZ_train.tif
│   │       │   ├── image1000_NZ_train.tif
│   │       │   └── image1001_NZ_train.tif
│   │       │       ... (5000 files total)
│   │       ├── 5000
│   │       │   ├── image5000_NZ_train.tif
│   │       │   ├── image5001_NZ_train.tif
│   │       │   └── image5002_NZ_train.tif
│   │       │       ... (5000 files total)
│   │       └── 10000
│   │           ├── image10000_NZ_train.tif
│   │           ├── image10001_NZ_train.tif
│   │           └── image10002_NZ_train.tif
│   │               ... (5000 files total)
│   │           ... (11 dirs total)
│   ├── val
│   │   ├── CH
│   │   │   └── 0
│   │   │       ├── image0_CH_val.tif
│   │   │       ├── image100_CH_val.tif
│   │   │       └── image101_CH_val.tif
│   │   │           ... (529 files total)
│   │   ├── NY
│   │   │   └── 0
│   │   │       ├── image0_NY_val.tif
│   │   │       ├── image100_NY_val.tif
│   │   │       └── image101_NY_val.tif
│   │   │           ... (529 files total)
│   │   └── NZ
│   │       └── 0
│   │           ├── image0_NZ_val.tif
│   │           ├── image100_NZ_val.tif
│   │           └── image101_NZ_val.tif
│   │               ... (529 files total)
│   └── test
│       ├── CH
│       │   ├── 0
│       │   │   ├── image0_CH_test.tif
│       │   │   ├── image1000_CH_test.tif
│       │   │   └── image1001_CH_test.tif
│       │   │       ... (5000 files total)
│       │   ├── 5000
│       │   │   ├── image5000_CH_test.tif
│       │   │   ├── image5001_CH_test.tif
│       │   │   └── image5002_CH_test.tif
│       │   │       ... (5000 files total)
│       │   └── 10000
│       │       ├── image10000_CH_test.tif
│       │       ├── image10001_CH_test.tif
│       │       └── image10002_CH_test.tif
│       │           ... (4400 files total)
│       ├── NY
│       │   ├── 0
│       │   │   ├── image0_NY_test.tif
│       │   │   ├── image1000_NY_test.tif
│       │   │   └── image1001_NY_test.tif
│       │   │       ... (5000 files total)
│       │   ├── 5000
│       │   │   ├── image5000_NY_test.tif
│       │   │   ├── image5001_NY_test.tif
│       │   │   └── image5002_NY_test.tif
│       │   │       ... (5000 files total)
│       │   └── 10000
│       │       ├── image10000_NY_test.tif
│       │       ├── image10001_NY_test.tif
│       │       └── image10002_NY_test.tif
│       │           ... (4400 files total)
│       └── NZ
│           ├── 0
│           │   ├── image0_NZ_test.tif
│           │   ├── image1000_NZ_test.tif
│           │   └── image1001_NZ_test.tif
│           │       ... (5000 files total)
│           ├── 5000
│           │   ├── image5000_NZ_test.tif
│           │   ├── image5001_NZ_test.tif
│           │   └── image5002_NZ_test.tif
│           │       ... (5000 files total)
│           └── 10000
│               ├── image10000_NZ_test.tif
│               ├── image10001_NZ_test.tif
│               └── image10002_NZ_test.tif
│                   ... (4400 files total)
├── lidar
│   ├── train
│   │   ├── CH
│   │   │   ├── 0
│   │   │   │   ├── lidar0_CH_train.copc.laz
│   │   │   │   ├── lidar1000_CH_train.copc.laz
│   │   │   │   └── lidar1001_CH_train.copc.laz
│   │   │   │       ... (5000 files total)
│   │   │   ├── 5000
│   │   │   │   ├── lidar5000_CH_train.copc.laz
│   │   │   │   ├── lidar5001_CH_train.copc.laz
│   │   │   │   └── lidar5002_CH_train.copc.laz
│   │   │   │       ... (5000 files total)
│   │   │   └── 10000
│   │   │       ├── lidar10000_CH_train.copc.laz
│   │   │       ├── lidar10001_CH_train.copc.laz
│   │   │       └── lidar10002_CH_train.copc.laz
│   │   │           ... (5000 files total)
│   │   │       ... (11 dirs total)
│   │   ├── NY
│   │   │   ├── 0
│   │   │   │   ├── lidar0_NY_train.copc.laz
│   │   │   │   ├── lidar10_NY_train.copc.laz
│   │   │   │   └── lidar1150_NY_train.copc.laz
│   │   │   │       ... (1071 files total)
│   │   │   ├── 5000
│   │   │   │   ├── lidar5060_NY_train.copc.laz
│   │   │   │   ├── lidar5061_NY_train.copc.laz
│   │   │   │   └── lidar5062_NY_train.copc.laz
│   │   │   │       ... (2235 files total)
│   │   │   └── 10000
│   │   │       ├── lidar10000_NY_train.copc.laz
│   │   │       ├── lidar10001_NY_train.copc.laz
│   │   │       └── lidar10002_NY_train.copc.laz
│   │   │           ... (4552 files total)
│   │   │       ... (11 dirs total)
│   │   └── NZ
│   │       ├── 0
│   │       │   ├── lidar0_NZ_train.copc.laz
│   │       │   ├── lidar1000_NZ_train.copc.laz
│   │       │   └── lidar1001_NZ_train.copc.laz
│   │       │       ... (5000 files total)
│   │       ├── 5000
│   │       │   ├── lidar5000_NZ_train.copc.laz
│   │       │   ├── lidar5001_NZ_train.copc.laz
│   │       │   └── lidar5002_NZ_train.copc.laz
│   │       │       ... (5000 files total)
│   │       └── 10000
│   │           ├── lidar10000_NZ_train.copc.laz
│   │           ├── lidar10001_NZ_train.copc.laz
│   │           └── lidar10002_NZ_train.copc.laz
│   │               ... (4999 files total)
│   │           ... (11 dirs total)
│   ├── val
│   │   ├── CH
│   │   │   └── 0
│   │   │       ├── lidar0_CH_val.copc.laz
│   │   │       ├── lidar100_CH_val.copc.laz
│   │   │       └── lidar101_CH_val.copc.laz
│   │   │           ... (529 files total)
│   │   ├── NY
│   │   │   └── 0
│   │   │       ├── lidar0_NY_val.copc.laz
│   │   │       ├── lidar100_NY_val.copc.laz
│   │   │       └── lidar101_NY_val.copc.laz
│   │   │           ... (529 files total)
│   │   └── NZ
│   │       └── 0
│   │           ├── lidar0_NZ_val.copc.laz
│   │           ├── lidar100_NZ_val.copc.laz
│   │           └── lidar101_NZ_val.copc.laz
│   │               ... (529 files total)
│   └── test
│       ├── CH
│       │   ├── 0
│       │   │   ├── lidar0_CH_test.copc.laz
│       │   │   ├── lidar1000_CH_test.copc.laz
│       │   │   └── lidar1001_CH_test.copc.laz
│       │   │       ... (5000 files total)
│       │   ├── 5000
│       │   │   ├── lidar5000_CH_test.copc.laz
│       │   │   ├── lidar5001_CH_test.copc.laz
│       │   │   └── lidar5002_CH_test.copc.laz
│       │   │       ... (5000 files total)
│       │   └── 10000
│       │       ├── lidar10000_CH_test.copc.laz
│       │       ├── lidar10001_CH_test.copc.laz
│       │       └── lidar10002_CH_test.copc.laz
│       │           ... (4400 files total)
│       ├── NY
│       │   ├── 0
│       │   │   ├── lidar0_NY_test.copc.laz
│       │   │   ├── lidar1000_NY_test.copc.laz
│       │   │   └── lidar1001_NY_test.copc.laz
│       │   │       ... (4964 files total)
│       │   ├── 5000
│       │   │   ├── lidar5000_NY_test.copc.laz
│       │   │   ├── lidar5001_NY_test.copc.laz
│       │   │   └── lidar5002_NY_test.copc.laz
│       │   │       ... (4953 files total)
│       │   └── 10000
│       │       ├── lidar10000_NY_test.copc.laz
│       │       ├── lidar10001_NY_test.copc.laz
│       │       └── lidar10002_NY_test.copc.laz
│       │           ... (4396 files total)
│       └── NZ
│           ├── 0
│           │   ├── lidar0_NZ_test.copc.laz
│           │   ├── lidar1000_NZ_test.copc.laz
│           │   └── lidar1001_NZ_test.copc.laz
│           │       ... (5000 files total)
│           ├── 5000
│           │   ├── lidar5000_NZ_test.copc.laz
│           │   ├── lidar5001_NZ_test.copc.laz
│           │   └── lidar5002_NZ_test.copc.laz
│           │       ... (5000 files total)
│           └── 10000
│               ├── lidar10000_NZ_test.copc.laz
│               ├── lidar10001_NZ_test.copc.laz
│               └── lidar10002_NZ_test.copc.laz
│                   ... (4400 files total)
└── ffl
    ├── train
    │   ├── CH
    │   │   ├── 0
    │   │   │   ├── image0_CH_train.pt
    │   │   │   ├── image1000_CH_train.pt
    │   │   │   └── image1001_CH_train.pt
    │   │   │       ... (5000 files total)
    │   │   ├── 5000
    │   │   │   ├── image5000_CH_train.pt
    │   │   │   ├── image5001_CH_train.pt
    │   │   │   └── image5002_CH_train.pt
    │   │   │       ... (5000 files total)
    │   │   └── 10000
    │   │       ├── image10000_CH_train.pt
    │   │       ├── image10001_CH_train.pt
    │   │       └── image10002_CH_train.pt
    │   │           ... (5000 files total)
    │   │       ... (11 dirs total)
    │   ├── NY
    │   │   ├── 0
    │   │   │   ├── image0_NY_train.pt
    │   │   │   ├── image1000_NY_train.pt
    │   │   │   └── image1001_NY_train.pt
    │   │   │       ... (5000 files total)
    │   │   ├── 5000
    │   │   │   ├── image5000_NY_train.pt
    │   │   │   ├── image5001_NY_train.pt
    │   │   │   └── image5002_NY_train.pt
    │   │   │       ... (5000 files total)
    │   │   └── 10000
    │   │       ├── image10000_NY_train.pt
    │   │       ├── image10001_NY_train.pt
    │   │       └── image10002_NY_train.pt
    │   │           ... (5000 files total)
    │   │       ... (11 dirs total)
    │   ├── NZ
    │   │   ├── 0
    │   │   │   ├── image0_NZ_train.pt
    │   │   │   ├── image1000_NZ_train.pt
    │   │   │   └── image1001_NZ_train.pt
    │   │   │       ... (5000 files total)
    │   │   ├── 5000
    │   │   │   ├── image5000_NZ_train.pt
    │   │   │   ├── image5001_NZ_train.pt
    │   │   │   └── image5002_NZ_train.pt
    │   │   │       ... (5000 files total)
    │   │   └── 10000
    │   │       ├── image10000_NZ_train.pt
    │   │       ├── image10001_NZ_train.pt
    │   │       └── image10002_NZ_train.pt
    │   │           ... (5000 files total)
    │   │       ... (11 dirs total)
    │   ├── processed-flag-all
    │   ├── processed-flag-CH
    │   └── processed-flag-NY
    │       ... (8 files total)
    ├── val
    │   ├── CH
    │   │   └── 0
    │   │       ├── image0_CH_val.pt
    │   │       ├── image100_CH_val.pt
    │   │       └── image101_CH_val.pt
    │   │           ... (529 files total)
    │   ├── NY
    │   │   └── 0
    │   │       ├── image0_NY_val.pt
    │   │       ├── image100_NY_val.pt
    │   │       └── image101_NY_val.pt
    │   │           ... (529 files total)
    │   ├── NZ
    │   │   └── 0
    │   │       ├── image0_NZ_val.pt
    │   │       ├── image100_NZ_val.pt
    │   │       └── image101_NZ_val.pt
    │   │           ... (529 files total)
    │   ├── processed-flag-all
    │   ├── processed-flag-CH
    │   └── processed-flag-NY
    │       ... (8 files total)
    └── test
        ├── CH
        │   ├── 0
        │   │   ├── image0_CH_test.pt
        │   │   ├── image1000_CH_test.pt
        │   │   └── image1001_CH_test.pt
        │   │       ... (5000 files total)
        │   ├── 5000
        │   │   ├── image5000_CH_test.pt
        │   │   ├── image5001_CH_test.pt
        │   │   └── image5002_CH_test.pt
        │   │       ... (5000 files total)
        │   └── 10000
        │       ├── image10000_CH_test.pt
        │       ├── image10001_CH_test.pt
        │       └── image10002_CH_test.pt
        │           ... (4400 files total)
        ├── NY
        │   ├── 0
        │   │   ├── image0_NY_test.pt
        │   │   ├── image1000_NY_test.pt
        │   │   └── image1001_NY_test.pt
        │   │       ... (5000 files total)
        │   ├── 5000
        │   │   ├── image5000_NY_test.pt
        │   │   ├── image5001_NY_test.pt
        │   │   └── image5002_NY_test.pt
        │   │       ... (5000 files total)
        │   └── 10000
        │       ├── image10000_NY_test.pt
        │       ├── image10001_NY_test.pt
        │       └── image10002_NY_test.pt
        │           ... (4400 files total)
        ├── NZ
        │   ├── 0
        │   │   ├── image0_NZ_test.pt
        │   │   ├── image1000_NZ_test.pt
        │   │   └── image1001_NZ_test.pt
        │   │       ... (5000 files total)
        │   ├── 5000
        │   │   ├── image5000_NZ_test.pt
        │   │   ├── image5001_NZ_test.pt
        │   │   └── image5002_NZ_test.pt
        │   │       ... (5000 files total)
        │   └── 10000
        │       ├── image10000_NZ_test.pt
        │       ├── image10001_NZ_test.pt
        │       └── image10002_NZ_test.pt
        │           ... (4400 files total)
        ├── processed-flag-all
        ├── processed-flag-CH
        └── processed-flag-NY
            ... (8 files total)

Pretrained model weights

Download

The recommended and fastest way to download the pretrained model weights is to run

python scripts/download_pretrained.py --model-root $MODEL_ROOT

Optionally you can also download the weights by running

git clone https://huggingface.co/rsi/PixelsPointsPolygons $MODEL_ROOT

Both options will download all checkpoints (as .pth) and results presented in the paper (as MS-COCO .json) into $MODEL_ROOT .

Code

Download

git clone https://github.com/raphaelsulzer/PixelsPointsPolygons

Installation

To create a conda environment named p3 and install the repository as a python package with all dependencies run

bash install.sh

or, if you want to manage the environment yourself run

pip install -r requirements-torch-cuda.txt
pip install .

⚠️ Warning: The implementation of the LiDAR point cloud encoder uses Open3D-ML. Currently, Open3D-ML officially only supports the PyTorch version specified in requirements-torch-cuda.txt.

Setup

The project supports hydra configuration which allows to modify any parameter either from a .yaml file or directly from the command line.

To setup the project structure we recommend to specify your $DATA_ROOT and $MODEL_ROOT in config/host/default.yaml.

To view all available configuration options run

python scripts/train.py --help

Predict demo tile

After downloading the model weights and setting up the code you can predict a demo tile by running

python scripts/predict_demo.py checkpoint=best_val_iou experiment=$MODEL_$MODALITY +image_file=demo_data/image0_CH_val.tif +lidar_file=demo_data/lidar0_CH_val.copc.laz

At least one of image_file or lidar_file has to be specified. $MODEL can be one of the following: ffl, hisup or p2p. $MODALITY can be image, lidar or fusion. The result will be stored in prediction.png.

Reproduce paper results

To reproduce the results from the paper you can run the following commands

python scripts/modality_ablation.py
python scripts/lidar_density_ablation.py
python scripts/all_countries.py

Custom training, prediction and evaluation

We recommend to first setup a custom experiment file $EXP_FILE in config/experiment/ following the structure of one of the existing files, e.g. ffl_fusion.yaml. You can then run

# train your model (on multiple GPUs)
torchrun --nproc_per_node=$NUM_GPU scripts/train.py experiment=$EXP_FILE

# predict the test set with your model (on multiple GPUs)
torchrun --nproc_per_node=$NUM_GPU scripts/predict.py experiment=$EXP_FILE evaluation=test checkpoint=best_val_iou

# evaluate your prediction of the test set
python scripts/evaluate.py experiment=$EXP_FILE evaluation=test checkpoint=best_val_iou

You could also continue training from a provided pretrained model with

# train your model (on a single GPU)
python scripts/train.py experiment=p2p_fusion checkpoint=latest

Citation

If you use our work please cite

@misc{sulzer2025p3datasetpixelspoints,
      title={The P$^3$ dataset: Pixels, Points and Polygons for Multimodal Building Vectorization}, 
      author={Raphael Sulzer and Liuyun Duan and Nicolas Girard and Florent Lafarge},
      year={2025},
      eprint={2505.15379},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.15379}, 
}

Acknowledgements

This repository benefits from the following open-source work. We thank the authors for their great work.

Datasets:

rsi
/

PixelsPointsPolygons

The P³ Dataset: Pixels, Points and Polygons
for Multimodal Building Vectorization

Raphael Sulzer^1,2 Liuyun Duan¹ Nicolas Girard¹ Florent Lafarge²

Abstract

Highlights

Dataset

Overview

Download

Structure

Pretrained model weights

Download

Code

Download

Installation

Setup

Predict demo tile

Reproduce paper results

Custom training, prediction and evaluation

Citation

Acknowledgements

Models trained or fine-tuned on rsi/PixelsPointsPolygons

rsi/PixelsPointsPolygons

The P3 Dataset: Pixels, Points and Polygons for Multimodal Building Vectorization

Raphael Sulzer1,2 Liuyun Duan1 Nicolas Girard1 Florent Lafarge2

Abstract

Highlights

Dataset

Overview

Download

Structure

Pretrained model weights

Download

Code

Download

Installation

Setup

Predict demo tile

Reproduce paper results

Custom training, prediction and evaluation

Citation

Acknowledgements

Models trained or fine-tuned on rsi/PixelsPointsPolygons

The P³ Dataset: Pixels, Points and Polygons
for Multimodal Building Vectorization

Raphael Sulzer^1,2 Liuyun Duan¹ Nicolas Girard¹ Florent Lafarge²