A 2D/3D multimodal data simulation approach with applications on urban semantic segmentation, building extraction and change detection

Abstract

Advances in remote sensing image processing techniques have further increased the demand for annotated datasets. However, preparing annotated multi-temporal 2D/3D multimodal data is especially challenging, both for the increased costs of the annotation step and the lack of multimodal acquisitions available on the same area. We introduce the Simulated Multimodal Aerial Remote Sensing (SMARS) dataset, a synthetic dataset aimed at the tasks of urban semantic segmentation, change detection, and building extraction, along with a description of the pipeline to generate them and the parameters required to set our rendering. Samples in the form of orthorectified photos, digital surface models and ground truth for all the tasks are provided. Unlike existing datasets, orthorectified images and digital surface models are derived from synthetic images using photogrammetry, yielding more realistic simulations of the data. The increased size of SMARS, compared to available datasets of this kind, facilitates both traditional and deep learning algorithms. Reported experiments from state-of-the-art algorithms on SMARS scenes yield satisfactory results, in line with our expectations. Both benefits of the SMARS datasets and constraints imposed by its use are discussed. Specifically, building detection on the SMARS-real Potsdam cross-domain test demonstrates the quality and the advantages of proposed synthetic data generation workflow. SMARS is published as an ISPRS benchmark dataset and can be downloaded from https://www2.isprs.org/commissions/comm1/wg8/benchmark_smars