Autoencoders for Semi-Supervised Water Level Modeling in Sewer Pipes with Sparse Labeled Data

More frequent and thorough inspection of sewer pipes has the potential to save billions in utilities. However, amount quality are impeded by an imprecise highly subjective manual process. It involves technicians judging stretches based on video from remote-controlled robots. Determining state these videos entails a great deal ambiguity. Furthermore, frequency with which different defects occur differs lot, leading imbalanced datasets. Such datasets represent poor basis for automating labeling process using supervised learning. With this paper we explore self-supervision as method reducing need large numbers well-balanced labels. First, our models learn data distribution more than million unlabeled images, then small number labeled examples used mapping learned representations relevant target variable, case, water level. We choose convolutional Autoencoder, Variational Autoencoder Vector-Quantised experiments. The best shown be classic Multi-Layer Perceptron achieving Mean Absolute Error 9.93. This is improvement 9.62 over fully baseline.