Deep features for skin detection and segmentation

This package contains the deep features extracted from the HGR (http://sun.aei.polsl.pl/~mkawulok/gestures/index.html) and ECU databases presenting hand gestures and human skin in color images, respectively. Additionally, a script (read_binary_file.py) that can be conveniently used for reading these features from the binary files is included in the package. The binary files with deep features are saved in the following form: N_features (4 bytes integer), N_labels (4 bytes integer), H (4 bytes integer), W (4 bytes integer), with H and W being the height and width of an input image. Then, there are H * W rows with N_features * feature_size (4 bytes float), and a label of a size label_size (4 bytes float) - each row represents a single pixel from an input image with the corresponding class label (the 0 label denotes the background, whereas 1 denotes skin). To extract features, we utilized a U-Net model trained in two different ways. In the Full variant, we train the model over the training data extracted from the ECU database presenting human skin color images, whereas in the TF variant, we exploit a model pre-trained for abnormality segmentation on a dataset of brain MRI volumes (kaggle.com/mateuszbuda/lgg-mri-segmentation) and available at https://pytorch.org/hub/mateuszbuda_brain-segmentation-pytorch_unet/, and fine tune it over the same training sample as in the Full variant. We dump the features extracted in the final convolutional layer of the model. The input images were rescaled with preserving the aspect ratio to 512x512 to match the original architecture (the background pixels are zeroed), and standardized using the mean and standard deviation of the original training dataset. Overall, we have 898 files containing the feature vectors extracted for all pixels in 898 HGR images (separately for color and grayscale variants, and separately for Full and TF), and 3998 files containing the feature vectors extracted for all pixels in 3998 ECU images (separately for color and grayscale variants, and separately for Full and TF). Important note: Due to the large size of the deep features, we include only a *sample* of deep features extracted for the HGR and ECU datasets here. To access full datasets, please contact Jakub Nalepa or Michal Kawulok.