Scan-Net: A Sparsely Encoded Convolutional Autoencoder for Semantic Segmentation of Unknown Terrain

A sparsely-encoded convolutional autoencoder architecture is proposed in this work for semantic segmentation of unknown terrain. The excellent feature extraction capabilities of the convolutional autoencoder (CAE) is utilized with the computation-efficient Echo State Network (ESN) for faster and efficient encoding, and semantic segmentation of unknown images. The proposed scheme manifests two CAEs trained with image and label data, and an ESN at the latent space of the two CAE to transform the encoded unknown image to semantic segmentation of different regions. The RUGD dataset of off-road images is used for training and validation of the proposed algorithm under variation of hyper-parameters. The proposed algorithm is implemented using Python and PyTorch, and simulation results demonstrate the effectiveness for semantic segmentation.