motokimura/spacenet6_solution

4th Place Solution for SpaceNet6 in CVPR'20 EarthVision Workshop

Jupyter Notebook

spacenet6_solution

motokimura's solution to SpaceNet6 challenge

Instructions for Final Scoring

See INSTRUCTION.md.

Sections below are only for the model development phase.

Please ignore the sections below in the final testing/scoring phase.

Instructions for Model Development

This section provides instructions for the model development phase.

Download SpaceNet6 data

# prepare data directory
DATA_DIR=${HOME}/data/spacenet6/spacenet6
mkdir -p ${DATA_DIR}
cd ${DATA_DIR}

# download and extract train data
aws s3 cp s3://spacenet-dataset/spacenet/SN6_buildings/tarballs/SN6_buildings_AOI_11_Rotterdam_train.tar.gz .
tar -xvf SN6_buildings_AOI_11_Rotterdam_train.tar.gz

# download and extract test data
aws s3 cp s3://spacenet-dataset/spacenet/SN6_buildings/tarballs/SN6_buildings_AOI_11_Rotterdam_test_public.tar.gz .
tar -xvf SN6_buildings_AOI_11_Rotterdam_test_public.tar.gz

Prepare training environment

# git clone source
PROJ_DIR=${HOME}/spacenet6_solution
git clone git@github.com:motokimura/spacenet6_solution.git ${PROJ_DIR}

# build docker image
cd ${PROJ_DIR}
./docker/build.sh

# launch docker container
ENV=desktop  # or "mac"
./docker/run.sh ${ENV}

Preprocess dataset

All commands below have to be executed inside the container.

./tools/compute_mean_std.py --image_subdir SAR-Intensity

./tools/compute_mean_std.py --image_subdir PS-RGBNIR

./tools/geojson_to_mask.py

./tools/split_dataset.py

./tools/separate_val_images.py

./tools/separate_val_labels.py

# optionally you can create AMI here

Train segmentation models

All commands below have to be executed inside the container.

EXP_ID=9999  # new experiment id
./tools/train_spacenet6_model.py [--config CONFIG_FILE] EXP_ID ${EXP_ID}

Test segmentation models

All commands below have to be executed inside the container.

EXP_ID=9999  # previous experiment id from which config and weight are loaded
./tools/test_spacenet6_model.py [--config CONFIG_FILE] --exp_id ${EXP_ID}

Ensemble segmentation models

All commands below have to be executed inside the container.

ENSEMBLE_EXP_IDS='[9999,9998,9997,9996,9995]'  # previous experiments used for ensemble
./tools/ensemble_models.py [--config CONFIG_FILE] ENSEMBLE_EXP_IDS ${ENSEMBLE_EXP_IDS}

Convert mask to polygon

All commands below have to be executed inside the container.

ENSEMBLE_EXP_IDS='[9999,9998,9997,9996,9995]'  # previous experiments used for ensemble
./tools/pred_array_to_poly.py [--config CONFIG_FILE] ENSEMBLE_EXP_IDS ${ENSEMBLE_EXP_IDS}

Test segmentation models (val images)

All commands below have to be executed inside the container.

EXP_ID=9999  # previous experiment id from which config and weight are loaded
./tools/test_spacenet6_model.py --config configs/test_to_val_images.yml --exp_id ${EXP_ID}

Ensemble segmentation models (val images)

All commands below have to be executed inside the container.

ENSEMBLE_EXP_IDS='[9999,9998,9997,9996,9995]'  # previous experiments used for ensemble
SPLIT_ID=0  # split id of the models above
TEST_IMAGE_DIR=/data/spacenet6/val_images/val_${SPLIT_ID}/SAR-Intensity
./tools/ensemble_models.py --config configs/test_to_val_images.yml ENSEMBLE_EXP_IDS ${ENSEMBLE_EXP_IDS} INPUT.TEST_IMAGE_DIR ${TEST_IMAGE_DIR}

Convert mask to polygon (val images)

All commands below have to be executed inside the container.

ENSEMBLE_EXP_IDS='[9999,9998,9997,9996,9995]'  # previous experiments used for ensemble
./tools/pred_array_to_poly.py --config configs/test_to_val_images.yml ENSEMBLE_EXP_IDS ${ENSEMBLE_EXP_IDS}

Train LGBM models

All commands below have to be executed inside the container.

./tools/train_lgbm.py --truth_csvs ${TRUTH_CSVS} --solution_csvs ${SOLUTION_CSVS} --imageids ${IMAGEIDS} --pred_dirs ${PRED_DIRS} --out_dir ${OUT_DIR}

Test LGBM models

All commands below have to be executed inside the container.

./tools/test_lgbm.py --solution_csv ${SOLUTION_CSV} --imageid ${IMAGEID} --pred_dir ${PRED_DIR} --models ${MODELS} --out ${OUT} [--iou_thresh ${IOU_THRESH}]