Using convolutional neural networks to pre-classify images for the humanitarian openstreetmap team (HOT & mapgive).
Python C++ Shell
Latest commit 4d19943 Sep 18, 2016 @larsroemheld committed on GitHub Update README.md
Permalink
Failed to load latest commit information.
caffe-segnet-patches Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
models Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
scripts (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
160314_Project_Report.pdf ... and added the project report. Mar 14, 2016
README.md Update README.md Sep 18, 2016
RunCaffeSegnet.py Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
checkFilesExist.py (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
convertMapImage.py (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
createCaffeData.py (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
exploreGeoJson.py explorations scripts for map data Mar 14, 2016
exploreMapImageLabels.py explorations scripts for map data Mar 14, 2016
getOSMmap.py (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
getTaskData.py (dirty) progress: get all data, create label files, split training se… Mar 12, 2016
parseLog.py Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
splitTrainVal.py Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
subtractMeanImage.py Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
test_segmentation_segnet.py Cleanup, add model specifications, add caffe segnet patches for inter… Mar 14, 2016
tileCoords.py Added project & task metadata crawler (esp. task status) Feb 19, 2016

README.md

OSM-HOT-ConvNet

Using convolutional neural networks to pre-classify images for the humanitarian openstreetmap team (HOT & mapgive).

This project uses satellite imagery to support map creation in the developing world. I gather my own dataset by downloading imagery released through the U.S. State Department’s MapGive project and using map data provided by the Humanitarian OpenStreetMap Team. Using this data, I train several Convolutional Neural Network models designed after the SegNet architecture to perform semantic image segmentation. The output of these models are map-like images that may in a later step be used to reconstruct map data, accelerating the work of online remote mapping volunteers. This paper details progress made towards this goal; my best model’s pixel-average test accuracy of about 69% does not allow production use yet. I conclude on notes for future work.

See the pdf-report for details.

I recommend this later report for more sophisticated approaches to classifying buildings specifically: https://devblogs.nvidia.com/parallelforall/exploring-spacenet-dataset-using-digits/