Models for Remote Sensing

Usage

Model training

1. Edit hyper-parameters in config.json or make your custom config file with the same format as config.json.
2. Run python train.py --config CONFIG_PATH if you're using your custom config file and CONFIG_PATH is the path to your config file. Without the --config flag, train.py will look for config.json in the mrs repo directory.

Model evaluation

1. Edit settings parameters (GPU ID, model path, etc.) in evaluate.py. Notice that DS_NAME should match the one in the config file of the trained model.
2. Run python evaluate.py.

Model inference

1. Edit settings parameters in infer.py in the same way as for model evaluation. DS_NAME should match the one in the trained model, too.
2. The FILE_LIST parameter takes the path to a .txt file which contains full paths of testing image files (one file path per row).
3. Run python infer.py.

Installation

Dependencies

• PyTorch >= 1.1
• TensorboardX
• TorchSummary
• Albumentation

Demos

• Histogram matching
• Object-wise scoring
• Results visualization

Encoders

Encoder Family	Encoder Name
VGG	vgg11, vgg11_bn, vgg13, vgg13_bn, vgg16, vgg16_bn, vgg19, vgg19_bn
ResNet	resnet18, resnet34, resnet50, resnet101
ResNeXt	resnext50_32x4d, resnext101_32x8d
WideResNet	wide_resnet50_2, wide_resnet101_2
Res2Net	res2net50_14w_8s, res2net50_26w_4s, res2net50_26w_6s, res2net50_26w_8s, res2net50_48w_2s, res2net101_26w_4s
Inception	inception_v3
SqueezeNet	squeezenet1_0, squeezenet1_1

Pretrained Models

Encoder Name	Decoder Name	Dataset	Label	Score (IoU)	Size	Model
VGG16	UNet	Inria	Building	78.56	207.3MB	Box
VGG19	UNet	Inria	Building	78.17	247.8MB	Box
ResNet34	UNet	Inria	Building	77.06	204.2MB	Box
ResNet50	UNet	Inria	Building	78.78	666.9MB	Box
ResNet101	UNet	Inria	Building	79.09	812.1MB	Box
VGG16	PSPNet	Inria	Building	76.23	171.1MB	Box
VGG19	PSPNet	Inria	Building	75.94	211.6MB	Box
ResNet34	PSPNet	Inria	Building	76.11	221.2MB	Box
ResNet50	PSPNet	Inria	Building	77.46	418.3MB	Box
ResNet101	PSPNet	Inria	Building	78.55	563.5MB	Box
ResNet34	DLinkNet	Inria	Building	75.67	248.5MB	Box
ResNet50	DLinkNet	Inria	Building	77.08	1.4GB	Box
ResNet50	UNet	DeepGlobe	Building	79.43	671.4MB	Box
ResNet101	UNet	DeepGlobe	Building	79.43	671.4MB	Box
ResNet101	DeepLabV3+	Inria	Building	79.17	464.6MB	Box
ResNet34	DLinkNet	DeepGlobe	Road	62.15	253MB	Box

Features

• Encoder Structures:
  • VGG
  • ResNet
  • DenseNet
  • SqueezeNet
  • InceptionNet
• Decoder Structures:
  • UNet
  • DLinkNet
  • PSPNet
  • DeepLabV3+
• Different Losses:
  • Xent
  • Jaccard Approximation
  • Focal Loss
  • Lovasz softmax (https://github.com/bermanmaxim/LovaszSoftmax/tree/master/pytorch)
  • Weighted combination of arbitrary supported losses
• Multi GPU Training
• Evaluation
  • Dataset Evaluator
  • Evaluate Report & Prediction Map
• Results visualization
• Class weights on criterions

Known Bugs

• Unable to do model-wise data parallel
• Failed to load models that trained on multiple gpus