OSMDeepOD - OSM and Deep Learning based Object Detection from Aerial Imagery
OSMDeepOD is a project about object detection from aerial imagery using open data from OpenStreetMap (OSM). The project uses the open source software library TensorFlow, with a retrained Inception V3 neuronal network.
This work started as part of a semester thesis autumn 2015 at Geometa Lab, University of Applied Sciences Rapperswil (HSR). See Twitter hashtag #OSMDeepOD for news.
- Paper AGIT 2016 (de) by S. Keller & S. Kurath, 7.7.2016
- Presentation (en) by S. Keller at GEOSmart Innovation Day 2016: tba.
- Presentation (en/de) by S. Bühler at PyDataZRH, 6.12.2016
The simplest way to use the detection process is to clone the repository and build/start the docker container.
git clone https://github.com/geometalab/OSMDeepOD.git cd OSMDeepOD/docker/ sudo docker build . -t osmdeepod sudo docker run -it --name osmdeepod -v ./:/objects osmdeepod bash
After the previous shell commands you have started a standalone instance of OSMDeepOD and you are connected to it. If you have a nvida GPU and nvidia-docker installed, you could use the "nvidia-docker" command to run the container for automatically usage of the GPU1.
To start the detection process use the src/role/main.py2 script.
- Use the manger option to select the detection area and start the detection with the --standalone parameter.
python3 main.py --config ./config.ini manager 9.345101 47.090794 9.355947 47.097288 --standalone
After the detection process has finished a "detected_nodes.json" file will appear with the results. If you like to use OSMDeepOD in a more parallel and distributed way have a look at the https://github.com/geometalab/OSMDeepOD-Visualize repository. There you have got the ability to use redis as a message queue and you can run many OSMDeepOD instances as workers.
The configuration works with an INI file. The file looks like the following:
[DETECTION] Network = /path/to/the/trained/convnet Labels = /path/to/the/label/file/of/the/convnet DetectionBarrier = 0.99 Word = crosswalk Key = highway Value = crossing ZoomLevel = 19 Compare = yes Orthofoto = other FollowStreets = yes StepWidth = 0.66 [REDIS] Server = 127.0.0.1 Port = 40001 Password = crosswalks BboxSize = 2000 Timeout = 5400
Some hints to the config file:
- "Word" is the key value of the labels file
- "Key" and "Value" builds the search Tag for OSM
- "Compare" means compared to OSM tagged Nodes
- "StepWidth" regulates the distance between the cut out images
- The section REDIS should be self explanatory, this is not necessary in the standalone mode
- "BboxSize" is the size in meters of the split large Bbox
- "Timeout" after the expired time the job does fail
To use your own Orthofotos you have to do the following steps:
- Add a new directory to
- Add a new module to the directory with the name:
- Create a class in the module with the name:
<Your_new_directory>Api(First letter needs to be uppercase)
- Implement the function
def get_image(self, bbox):and returns a pillow image of the bbox
- After that you can use your api with the parameter
If you have problems with the implementation have a look at the wms or other example.
During this work, we have collected our own dataset with swiss crosswalks and non-crosswalks. The pictures have a size of 50x50 pixels and are available by request.
Picture 3: Crosswalk Examples
Picture 4: No Crosswalk Examples
At the moment, we support python 3.5
In order to use volumes, I recommend using docker >= 1.9.x
Bounding Box of area to analyze
To start the extraction of crosswalks within a given area, the bounding box of this area is required as arguments for the manager. To get the bounding box the desired area, you can use https://www.openstreetmap.org/export to select the area and copy paste the corresponding coordinates. Use the values in the following order when used as positional arguments to manager:
left bottom right top
- 1: The crosswalk_detection container is based on the nvidia/cuda:7.5-cudnn4-devel-ubuntu14.04 image, may you have to change the base image for your GPU. ↩
- 2: For more information about the main.py use the -h option. ↩
Big Data; Data Science; Data Engineering; Machine Learning; Artificial Intelligence; Neuronal Nets; Imagery; Volunteered Geographic Information; Crowdsourcing; Geographic Information Systems; Infrastructure; Parallel Programming.