Setting up the repository:

After cloning the repository and installing the (Nvidia) Docker:

$ docker build -f ./docker/DockerfileCUDA11 --rm -t rooftop .

$ docker run --ipc=host --name rooftopml --gpus all -it --network host -v "$PWD:/home/docker/repository" rooftop /bin/bash

If you are using a machine without Nvidia GPU, remove the --gpus all parameter.

Enter container to run some experiments.

docker start -i rooftopml

Image tiling:

python tools/create_images_dataset.py -id {path to images} -ad {path to masks} -od {output directory} -ts 512 512 -st 512 512 -sc 0.3 -fmt .png -w 12

Arguments:

• -ts size of tiles (x, y)
• -st stride (x, y)
• -sc Scale factor
• -fmt output image format
• -w number of workers to be spawned for tiling

Building models on AIRS dataset

Deep Watershed Transform (DWT) segmentation model:

python tools/train.py -c configs/airs_pretrain_unet_dwt.yaml

To be able to use mmdetection library, please refer to the original documentation of the library to install the package inside the container.

Generating COCO annotations (train/val/test)

python tools/coco_annotations.py -i data/airs_proto/train -o data/airs_proto/train_annotations.json --minimize

python tools/coco_annotations.py -i data/airs_proto/val -o data/airs_proto/val_annotations.json --minimize

python tools/coco_annotations.py -i data/airs_proto/test -o data/airs_proto/test_annotations.json --minimize

Arguments:

• -id Directory containing dataset
• -od JSON file to write to
• -m Whether to minimize the output json file or not

Training Point Rend

python mmdetection/tools/train.py configs/pointrend_r50.py

Inferencing a folder of GeoTiff images with DWT model

python tools/inference_geotiff.py -c configs/airs_pretrain_unet_dwt_inference.yaml -id data/inference_examples/