MRPolyBuild

Official repository for paper "Rethinking Resolution: Large-Scale Polygonal Building Detection Using Medium-Resolution (3-5m) Satellite Data".

News

• 2026-08: Released training/test annotations, official checkpoints, evaluation code and a single-image inference demo.

Downloads

All released assets are hosted in the MRPolyBuild Google Drive folder.

Checksums of every file are provided in SHA256SUMS.txt at the folder root.

Annotations (satellite imagery is not redistributed)

File	Description	Link
data/mrpolybuild_merged_ann.tar.zst (~15 GB)	162,531 per-tile building annotations fused from Microsoft, Google and CLSM datasets	link
data/test_6k_osm.tar.zst (~909 MB)	OSM annotations for the 6k test tiles	link
data/train_roi.geojson	Training RoI polygons	link
data/test_roi.geojson	Test RoI polygons	link
data/test_6k_metrics.tar.zst	Evaluation metric summaries	link
data/coco_no_ann.json	Test image list (COCO format)	link

Checkpoints

Model	File	Size	Link
Stage-1 semantic segmentation (50 epochs)	checkpoints/seg-based-det_8x_multi-source_convnext-v2-b_50e_planet_basemap_global/epoch_50.pth	2.0 GB	link
Stage-2 polygon refinement (320k iterations)	checkpoints/gcp_ins-v2_8x_right-ang-v2_seg-based-det_convnext-v2-b_320k_planet_basemap_global/iter_320000.pth	1.1 GB	link

Demo

File	Link
Predicted GeoJSON for the demo tile	link
Sample PlanetScope tile (Saint-Martin-de-Crau, France)	downloaded & cropped by bash demo/download_demo_data.sh from the ESA public sample (not redistributed here)

Installation

First, create a conda environment:

conda env create -f environment.yaml -n MRPolyBuild

Then install the package:

pip install -e .

Extra dependencies required for evaluation:

pip install -r requirements/polygon.txt

Data Preparation

The satellite images used in this work are from PlanetScope company, which are commercial data. Hence we are not allowed to publish them here. We do provide the annotation files and RoI files that can be used to search the corresponding tiles from PlanetScope imagery, and building polygon results of our method and comparison methods that can be used to reproduce the evaluation metrics in the paper.

Download

Download the annotation files from the Drive folder (or use the per-file links above) and extract them:

mkdir -p data
tar --use-compress-program=unzstd -xf mrpolybuild_merged_ann.tar.zst -C data
tar --use-compress-program=unzstd -xf test_6k_osm.tar.zst -C data

The resulting directory structure is:

data/
├── train_roi.geojson
├── test_roi.geojson
├── train_160k/
│   ├── img/                  # not provided, see below
│   ├── merged_ann/           # 162,531 per-tile annotation jsons
│   └── ...
└── test_6k/
    ├── img/                  # not provided, see below
    ├── osm/geojson/          # test annotations (GeoJSON)
    ├── osm/json/             # test annotations (JSON)
    └── ...

train_roi.geojson and test_roi.geojson are lists of bounding box RoIs that define the sampling regions of the training and testing data, and can be used to download the corresponding tiles from publicly available data sources. data/train_160k/merged_ann contains the building annotations fused from the Microsoft, Google and CLSM datasets, one JSON file per tile, with the filename matching the tile image filename.

No satellite images are provided. To train or test with your own networks, prepare the image tiles in .tif format and place them in img/, making sure the image and annotation filenames are paired.

Checkpoints

Download the checkpoints and place them under checkpoints/, keeping the directory names shown in the table above, e.g.:

checkpoints/gcp_ins-v2_8x_right-ang-v2_seg-based-det_convnext-v2-b_320k_planet_basemap_global/iter_320000.pth

The Stage-2 checkpoint is an end-to-end model (backbone + segmentation head + polygon refinement head). When training the Stage-2 model yourself, set the load_from variable in the Stage-2 config to the Stage-1 checkpoint path.

Evaluation

To reproduce the evaluation metrics in the paper, run:

python tools/eval_planet_metrics_by_geojson_list.py \
    --product-name gcp_ins-v2_8x_right-ang-v2_320k \
    --pred-geojson-pattern "data/test_6k/{product_name}/geojson/*.geojson" \
    --gt-geojson-pattern "data/test_6k/osm/geojson/*.geojson" \
    --out-base-path "data/test_6k/metrics/{product_name}"

Notes:

• --product-name can be passed multiple times to evaluate several prediction products in one run.
• The Globe filter is evaluated by default; continent/country filters can be enabled inside the script.
• Results are written to overall_Globe.json and summary_table.txt in the output directory.
• --match-mode selects whether to pair predictions with ground truth by common files (common) or by all ground-truth files (gt_driven).

Train

Training MRPolyBuild includes the following two steps.

Train Semantic Segmentation Networks

In the first stage, we train a plain semantic segmentation network by running:

python tools/train.py configs/planet_basemap/seg-based-det_8x_multi-source_convnext-v2-b_50e_planet_basemap_global.py

You will need to change the data_root in configs/_base_/datasets/planet_basemap_single_ann_2023q2_global_8x_train-160k.py to your data directory.

Train Polygon Refinement Networks

In the second stage, we train the polyline refinement module by running:

python tools/train.py configs/planet_basemap/gcp_ins-v2_8x_right-ang-v2_seg-based-det_convnext-v2-b_320k_planet_basemap_global.py

Make sure you have specified the load_from variable to the network weights achieved in the first stage.

Inference

Quick Start — Single-Image Inference

1. Download the Stage-2 checkpoint and place it under checkpoints/ (see Checkpoints).
2. Download and crop the demo tile (original image from the ESA public PlanetScope sample):

bash demo/download_demo_data.sh

3. Run the inference:

python tools/inference_planet_basemap.py demo/configs/inf_demo.py

The demo config processes demo/data/saint_martin_de_crau_town.tif, a 2048x2048 PlanetScope visual tile (Saint-Martin-de-Crau, France), with the official 8x upsampled inference pipeline (256x256 crops upsampled 8x to 2048x2048, assembled at 8x output resolution).

Input	MRPolyBuild prediction (5,015 buildings)

Input / prediction side-by-side and a 4x zoom of the town centre:

The predicted polygons are saved as GeoJSON at:

demo/output/gcp_ins-v2_8x_right-ang-v2_320k_demo/geojson/saint_martin_de_crau_town.geojson

Inference with PlanetScope Satellite Images

To run inference on large satellite images, specify a folder containing .tif files. Refer to the configuration file configs/planet_basemap/inf_gcp_ins-v2_8x_right-v2_convnext-v2-b_320k_planet_basemap_filtered_oceania.py for details.

After the data path is configured, the inference pipeline can be run using:

python tools/inference_planet_basemap.py configs/planet_basemap/inf_gcp_ins-v2_8x_right-v2_convnext-v2-b_320k_planet_basemap_filtered_oceania.py

Inference with Existing Probability Maps

We provide a model that can perform zero-shot building polygonal mapping from building probability maps (usually generated by a neural network).

Here we provide an example of converting building probability maps from the Google 2.5D Temporal dataset to polygonized buildings:

python tools/inference_planet_basemap_from_tif_probs.py configs/planet_basemap/inf_gcp_ins-v2_8x_right-v2_convnext-v2-b_320k_planet_basemap_google25d.py

You may want to configure the paths in the save_cfg variable in the configuration file.