Randomly connected neural networks for self-supervised monocular depth estimation

Abstract

A plethora of deep learning-based methods have been proposed for self-supervised monocular depth estimation. The majority of these models utilise a U-Net-based architecture for disparity estimation. However, this architecture may not be optimal, and as shown in self-supervised approaches, replacing standard U-Net encoder with more complex architectures like the ResNet encoder achieves superior performance. In monocular depth estimation, the success of the method is attributed to the model architecture design capable of extracting high order features from a single image and the loss function for enforcing a strong feature distribution. In this work, a novel randomly connected encoder-decoder architecture has been designed for self-supervised monocular depth estimation. To enable efficient search in the connection space, the ‘cascaded random search’ approach is introduced for the generation of random network architectures. We introduce a new U-Net topology capable of utilising the semantic information of feature maps. For high quality image reconstructions, a loss function has been proposed, which efficiently extends perceptual and adversarial loss to multiple scales. We conduct performance evaluation on two surgical datasets, including comparisons to state-of-the-art self-supervised depth estimation methods. The performance evaluation analysis verifies the superiority of our randomly connected network architecture.