MUNO21 is a dataset and benchmark for machine learning methods that automatically update and maintain digital street map datasets. Previous datasets focus on road extraction, and measure how well a method can infer a road network from aerial or satellite imagery. In contrast, MUNO21 measures how well a method can modify the road network data in an existing digital map dataset to make it reflect the latest physical road network visible from imagery. This task is more practical, since it doesn't throw away the existing map, but also more challenging, as physical roads may be constructed, bulldozed, or otherwise modified.
For more details, see https://favyen.com/muno21/.
This repository contains the code that was used to create MUNO21, as well as code for working with the dataset and computing evaluation metrics.
Compiler and application requirements include the following. The versions are what we use and older versions make work as well.
- Go 1.16+ (with older versions, module-aware mode must be enabled)
- Python 3.5
- osmium-tool 2.16.0 (only needed for dataset pre-processing)
- ImageMagick 6.8 (only needed for dataset pre-processing)
Python requirements are in requirements.txt, and can be installed with:
pip install -r requirements.txt
These requirements should be sufficient to run dataset pre-processing, automatic candidate generation and clustering, visualization, metric evaluation, and post-processing with removing G_extra and fusing new roads into the base map.
To run the included map update methods, a range of additional requirements are needed, depending on the particular method:
- TensorFlow 1.15 (not 2.0)
- pytorch 1.7
- scipy 1.4
- OpenCV
- rdp
Download and extract the MUNO21 dataset:
wget https://favyen.com/files/muno21.zip
unzip muno21.zip
In the commands below, we may assume that you have placed the dataset in /data/:
mv mapupdate/ /data/
The dataset includes aerial image and road network data in large tiles around several cities, along with annotations that specify the map update scenarios. Some steps below will require road network data to be extracted in windows corresponding to the scenarios:
cd muno21/go/
mkdir /data/identity
export PYTHONPATH=../python/
python ../methods/identity/run.py /data/graphs/graphs/ /data/annotations.json /data/identity/
Aerial imagery is available as JPEG images in the naip/jpg/ folder.
These images are obtained from NAIP.
Road networks are available as .graph files in the graphs/graphs folder.
See https://favyen.com/muno21/graph-format.txt for a description of the data
format of these files. Note that, in contrast to some other datasets, road
networks are NOT represented as images -- instead, they are undirected spatial
networks, where vertices are labeled with (x,y) coordinates and edges
correspond to road segments. The (x,y) coordinates indicate pixels in the
corresponding JPEG image.
Note that two versions of the road network are available in this format.
- {region_x_y_time}.graph: only includes public roads suitable for motor vehicles.
- {region_x_y_time}_all.graph: includes most other "ways" that appear in OpenStreetMap.
The original OpenStreetMap data is available in the graphs/osm/ folder, in
files encoded under the OSM PBF format. Methods may take advantage of the
additional information in these files, such as various road attributes. To
convert longitude-latitude coordinates to pixel coordinates, see go/lib/regions.go
and go/preprocess/osm_to_graph.go.
The MUNO21 dataset includes 1,294 map update scenarios. Each scenario specifies a pre-change timestamp, post-change timestamp, and a bounding box window where some change occurred.
The input is aerial imagery from each of four years, along with road network data from a specific pre-change year (usually 2012 or 2013).
The ground truth label is the road network from a specific post-change year (usually 2018 or 2019) inside the bounding box window.
During training, a method may use all aerial imagery and road network data from
the training regions (see train.json). To facilitate self-supervised learning,
methods may also use all aerial imagery in the test regions (see test.json),
but only road network data from 2012 or 2013 in those regions.
During inference, for a given scenario, a method has access to the same data that is available during training. It additionally has access to road network data from all regions at the pre-change timestamp, although since this is usually 2012 or 2013, this usually does not actually provide any more data.
The method should output a road network corresponding to the physical roads visible in the aerial imagery at the post-change timestamp inside the bounding box window.
Methods are compared in terms of their precision-recall curves.
Recall measures how much closer the output road networks are to the ground truth data (post-change road network) than the pre-change road networks. Two alternative ways of comparing road networks, PixelF1 and APLS, are used.
Precision measures how frequently a method makes incorrect modifications to the road network in scenarios where no change has occurred between the pre- and post-change timestamps.
A method may expose a single real-valued parameter that provides a tradeoff between precision and recall. For example, a method that infers road networks using image segmentation may expose the segmentation probability confidence threshold for the "road" class as a parameter -- increasing this threshold generally provides higher precision but lower recall. Methods are compared in terms of their precision-recall curves when varying this parameter.
Scenarios are specified in the annotations.json file. Let annotation refer
to one annotation JSON object.
Each scenario specifies a spatial window in pixel coordinates where the map has
changed: annotation['Cluster']['Window']. A method may use imagery and road
network data outside that window, but its output road network should span that
window plus 128-pixel padding; it will be evaluated only inside the window (with
no padding), but the padding ensures that the evaluation metrics are computed
correctly along the boundary of the window.
Currently, the pre-change timestamp is always 2013, and the post-change timestamp is always the year of the most recent aerial image (either 2018 or 2019).
Refer to the documentation in methods/{classify,recurrentunet,road_connectivity,roadtracerpp,sat2graph}.
Each method besides classify is taken from a publicly available
implementation (see README in each method directory.) We make minor changes to
make them work with MUNO21. We also find many bugs in road_connectivity which
we have to manually fix, and we adapt Sat2Graph to work with Python3.
road_connectivity and recurrentunet will only work with Python 2.7.
Applying a method to infer road networks should yield a directory containing subdirectories (corresponding to different confidence thresholds) that each contain .graph files. Most methods require post-processing under our map fusion approach before evaluation.
Suppose that you have computed the outputs of MAiD in /data/maid/out/.
Then, for each confidence threshold:
mkdir /data/maid/fuse/
mkdir /data/maid/fuse/10/
go run postprocess/fuse.go /data/annotations.json normal /data/identity/ /data/maid/out/10/ /data/maid/fuse/10/
Optionally, visualize an inferred road network. Below, 6 can be changed to any
annotation index corresponding to /data/annotations.json.
go run vis/visualize_inferred.go /data/annotations.json 6 /data/naip/jpg/ /data/graphs/graphs/ /data/maid/fuse/10/ default ./
The command above should produce an image ./6.jpg.
For each confidence threshold, run e.g.:
python metrics/apls.py /data/annotations.json /data/maid/fuse/10/ /data/graphs/graphs/ /data/test.json
go run metrics/geo.go /data/annotations.json /data/maid/fuse/10/ /data/graphs/graphs/ /data/test.json
Above, the first command computes APLS (which takes a long time to run) while
the second computes PixelF1 (aka GEO metric). These commands produce
scores.json and geo.json files respectively in the /data/maid/fuse/10/
directory containing metric outputs for each test scenario.
To obtain error rate:
go run metrics/error_rate.go /data/annotations.json /data/maid/fuse/10/ /data/graphs/graphs/ /data/test.json
To produce a precision-recall curve from the scores across multiple confidence thresholds, run:
python metrics/score_details.py /data/annotations.json /data/maid/fuse/{10,20,30,40,50}/geo.json
The documentation below outlines how the dataset was built. You do not need to follow these steps unless you are trying to replicate the dataset from raw NAIP aerial images from Google EarthEngine and OpenStreetMap history dumps.
We preprocess raw NAIP and OSM data using the code in go/preprocess.
- Obtain NAIP images from Google EarthEngine.
- Obtain us-internal.osh.pbf from https://download.geofabrik.de/north-america/us.html
- Extract history around individual cities:
go run preprocess/osm_space_filter.go /data/graphs/big/us-internal.osh.pbf /data/graphs/history/ - Extract OSM dumps at different times:
python3 preprocess/osm_time_filter.py /data/graphs/history/ /data/graphs/osm/ - Convert NAIP images to JPG:
python3 preprocess/tif_to_jpg.py /data/naip/tif/ /data/naip/jpg/ - Record the NAIP image sizes (needed for coordinate transforms and such):
python3 preprocess/save_image_sizes.py /data/naip/jpg/ /data/sizes.json - Convert to MUNO21 .graph file format:
go run preprocess/osm_to_graph.go /data/graphs/osm/ /data/graphs/graphs/ - Randomly split the cities into train/test:
python3 preprocess/pick_train_test.py /data/graphs/history/ /data/ - (Optional) Visualize the graph and image extracted at a tile:
python3 vis/vis.py /data/naip/jpg/ny_1_0_2019.jpg /data/graphs/graphs/ny_1_0_2018-07-01.graph out.jpg
We then generate and cluster candidates.
- Candidate generation:
go run annotate/find_changed_roads.go /data/graphs/graphs/ /data/changes/ - Clustering:
go run annotate/cluster_changes.go /data/changes/ /data/cluster/ - No-change windows:
go run annotate/find_nochange.go /data/graphs/graphs/ /data/cluster-nochange/ - Output visualizations for annotation:
go run annotate/visualize_clusters.go /data/cluster/ /data/naip/jpg/ /data/graphs/graphs/ /data/vis/
After using the annotation tools like go/annotate, we process the output annotations into JSON file:
- Convert annotation data to JSON:
go run process_annotations.go /data/cluster/ /data/annotations.txt /data/cluster-nochange/ /data/annotations.json