Abstract

In this work we explore two different scenarios of 3D reconstruction. The first, urban scenes, is approached using a deep learning network trained to identify structurally important classes within aerial imagery of cities. The network was trained using data taken from ISPRS benchmark dataset of the city of Vaihingen. Using the segmented maps generated by the network we can proceed to more accurately reconstruct the scenes by a process of clustering and then class specific model generation. The second scenario is that of underwater scenes. We use two separate networks to first identify caustics and then remove them from a scene. Data was generated synthetically as real world datasets for this subject are extremely hard to produce. Using the generated caustic free image we can then reconstruct the scene with more precision and accuracy through a process of structure from motion. We investigate different deep learning architectures and parameters for both scenarios. Our results are evaluated to be efficient and effective by comparing them with online benchmarks and alternative reconstruction attempts. We conclude by discussing the limitations of problem specific datasets and our potential research into the generation of datasets through the use of Generative-Adverserial-Networks.