The functionality of sewer pipes is important in order to avoid social problems and the spread of diseases. For that
purpose, the project tackles the research question “Is it feasible to detect defects in sewer pipes using deep-learning with convolutional neural networks”. Taking a dataset of text-annotated videos and annotating them by hand, 1000 samples of infiltration defect were finalized. Splitting 60% train, 20% validation and 20% testing for experiments using two networks – U-Net and Mask-RCNN. After training the networks and performing validation, U-Net resulted in 0.85 IoU score, while Mask-RCNN 0.71. Hence it is concluded that detection of infiltration in sewer pipes is possible with the use of U-Net, however, more research needs to be done with a more reliable dataset and annotations in order for this to be applicable in the real world.