I’m happy to announce the RoboSat v1.1.0 release — our open source end-to-end pipeline for feature extraction from aerial and satellite imagery.
This release is powered by major community contributions from features like fine-tuning trained models, to training and prediction speedups, and bug fixes! Thank you to our contributors!
Here is an overview of what you will find in the v1.1.0 release
rs train: new --checkpoint argument to re-start training (fine-tune)
from a trained model checkpoint. Thanks https://github.com/ocourtin
rs train: memory usage reduction during validation by disabling
expensive gradient computation. Thanks https://github.com/Jesse-jApps
rs train, rs predict: speedups using multiple workers and
doing metric calculation on GPU. Thanks https://github.com/ocourtin
rs merge: polygon orientation fixes to respect the GeoJSON
specification (right-hand rule). Thanks https://github.com/marsbroshok
You can find automatically built Docker images at https://hub.docker.com/r/mapbox/robosat/
For our next release we already have a couple amazing community contributions
rs extract: a new road extraction handler to automatically
generate a training dataset for roads; Thanks https://github.com/DragonEmperorG
New state of the art losses (like the Lovasz loss) and improved metrics among other improvements. Thanks https://github.com/ocourtin
Here’s why you should be excited for the Lovasz loss: the following shows a preview
Left to right: image / label / cross entropy / miou / lovasz
In contrast to the current default cross entropy loss the Lovasz loss will be an amazing default for robosat’s binary models and greatly improve common feature extraction use-cases!
Again, thanks for these amazing community contributions!
As usual hit us up for feedback be it at OSM hack weekends, on tickets, IRC, or the OSMUS Slack. Also check out previous diary posts about robosat: