- data_loader: where pytorch dataset is defined for this framework
- data_utils: some useful functions for creating & managing datasets
It is recommended to use the mean and variance calculated for each dataset in the table below to for the configuration file. An alternative way is to use the mean and std of the ImageNet dataset.
The function (get_ds_stats()) to calculate mean and std of given list of images is defined in data_utils. Either use dataset specific mean&std or not, it is important to have the same set of mean&std in both training and evaluation.
| Dataset Name | Label | Web Page | Mean | Std |
|---|---|---|---|---|
| Inria | Building | Link | (0.41776216, 0.43993309, 0.39138562) | (0.18476704, 0.16793099, 0.15915148) |
| DeepGlobeBuilding | Building | Link | (0.34391829, 0.41294382, 0.45617783) | (0.10493991, 0.09446405, 0.08782307) |
| DeepGlobeRoad | Road | Link | (0.40994515, 0.38314009, 0.28864455) | (0.12889884, 0.10563929, 0.09726452) |
| DeepGlobeLand | Urban land, Agriculture land, Rangeland, Forest land, Water, Barren land | Link | (0.38475783 0.34792641 0.24514727) | (0.12838193 0.10063904 0.08192587) |
| MIT Road | Road | Link | (0.4251811, 0.42812928, 0.39143909) | (0.22423858, 0.21664895, 0.22102307) |