omranlm / fast_segmentation

Building segmentation hackathon

Visualizer links

• http://13.95.159.165/tool/

Setup

conda config --set channel_priority strict
conda env create --file environment.yml
conda activate torchgeo

Pipeline steps

• Upload geojson labels to data/labels/
• Convert labels from EPSG:4326 (lat/lon) to the coordinate system (CRS) of the imagery, in this case, EPSG:32616
  • ogr2ogr -f GeoJSON -t_srs EPSG:32616 demo_annotations_epsg32616.geojson demo_annotations.geojson
• Create masks
  • python create_mask_from_annotations.py --input-fn data/labels/demo_annotations_epsg32616.geojson --target-fn data/imagery/16_pre_imagery_cropped.tif --output-dir data/masks/ --overwrite
  • NOTE: create_mask_from_annotations.py will need to be edited with the class names used in the web-tool
• Buffer the masks
  • Run Apply distance buffer to mask.ipynb
• Train models
  • Run Train.ipynb
• Inference
  • python inference.py --input-model-checkpoint output/runs/unet-resnet18-imagenet-lr_0.001/last.ckpt --input-image-fn data/imagery/16_pre_imagery_cropped.tif --output-dir predictions/unet-resnet18-imagenet-lr_0.001/ --overwrite --gpu 1