SpaceNet6 Challenge

Models

Unet models with different encoders from segmentation models libruary. All base models used were pre-trained on Imagenet dataset.

Dataset

The data were provided by th challenge orginisers. The script to download the dataset is in scripts/download_spacenet6_data. The corresponding masks were uploaded on kaggle, the donload scrips is in scripts/download_masks.

Images preprocessing and augmentations

The original images were scaled to 512 x 512 px resolution. First, the models were pre-trained on RGB data. For the RGB images I added the grayscale to have 4 input channes for the pre-train, similar to SAR data. The RGB dataset is in src/datasets/spacenet_rgb.py. The data were normalised as in ImageNet.

Fig. 1 An example of RGB and grayscale data and the corresponding binary mask.

The images were agmented using albumentations libruary. The dictionary of possible transforms is in src/datasets/transforms.py

Fig. 2 Augmented RGB and grayscale data with the corresponding binary mask.

After the pre-train models were trained on 4-channel SAR data. The raw data were normalised using 98% percintile and then as in ImageNet. The SAR dataset is in src/datasets/spacenet.py

Fig. 3 4 channels SAR data with the corresponding binary mask.

The transforms "medium" and "d4" were used for training, ""flip_bright"" for validation, see transforms in src/datasets/transforms.py

Fig. 4 Augmented 4 channels SAR data.

Prepare environment

1. Install anaconda
2. Run scripts/create_env.sh bash file to set up the conda environment

Training

Training script is in src/pretrain_runner.py for RGB data and in src/train_runner.py for SAR data, respectively.

Set up your own path ways in src/configs.py. Basic model: Unet Se-ResNext101 Set IMS_SIZE in configs for transforms.

Make folds

python -m src.folds.make_folds

Create empty masks for images with no objects

python -m src.utils.make_empty_masks

Test the train runners

python -m src.pretrain_runner --model-name "unet_se_resnext101_32x4d" --encoder "se_resnext101_32x4d" --debug True --image-size 224 --epochs 2 --lr 1e-3 --batch-size 16 --num-workers 2

python -m src.train_runner --model-name "unet_se_resnext101_32x4d" --encoder "se_resnext101_32x4d" --debug True --image-size 224 --epochs 2 --lr 1e-3 --batch-size 16 --num-workers 2

Pretrain the model on RGB + grayscale images

nuhop python -m src.pretrain_runner --model-name "unet_se_resnext101_32x4d" --encoder "se_resnext101_32x4d" --image-size 512 --epochs 200 --lr 1e-3 --batch-size 6 --num-workers 4 & tail -f nohup.out

Train the model on SAR data

You may use the model pretrained on RGB + grayscale images; set the --checkpoint argument nohup python -m src.train_runner --model-name "unet_se_resnext101_32x4d" --encoder "se_resnext101_32x4d" --image-size 512 --checkpoint = '<pretrained_rgb_model>' --epochs 200 --lr 1e-3 --batch-size 6 --num-workers 4 & > nopup2.out tail -f nohup.out