SynRS3D: A Synthetic Dataset for Global 3D Semantic Understanding from Monocular Remote Sensing Imagery Neural Information Processing Systems (Spotlight), 2024 For more details, please refer to our paper and visit our GitHub repository. Overview TL;DR:SynRS3D is a comprehensive synthetic remote sensing dataset designed to improve global 3D semantic understanding from monocular high-resolution imagery. It includes data for three key tasks: Height estimation Land cover mapping Building change detection Dataset Structure The dataset consists of 17 folders and includes a total of 69,667 images at a resolution of 512x512. After downloading and extracting the files, ensure the directory structure follows this format: ${DATASET_ROOT} # Example: /home/username/project/SynRS3D/data/grid_g05_mid_v1├── opt # RGB images (.tif), also used as post-event images for building change detection├── pre_opt # RGB images (.tif), used as pre-event images for building change detection├── gt_nDSM # Normalized Digital Surface Model (nDSM) images (.tif)├── gt_ss_mask # Land cover mapping labels (.tif)├── gt_cd_mask # Building change detection masks (.tif, 0 = no change, 255 = change area)└── train.txt # List of training data filenames The land cover mapping labels (`gt_ss_mask`) are mapped to the following categories: Bareland: 1 Rangeland: 2 Developed Space: 3 Road: 4 Trees: 5 Water: 6 Agriculture land: 7 Buildings: 8 Image Breakdown by Folder The dataset is organized into grid-like and irregular terrain. It includes a range of ground sampling distances (GSDs) and variations in building heights. The folder naming convention indicates these characteristics: - `grid` = grid-like terrain - `terrain` = irregular terrain - `g005`, `g05`, `g1` = GSD ranges (0.05m–0.3m, 0.3m–0.6m, and 0.6m–1m, respectively) - `low`, `mid`, `high` = building height variations The dataset includes the following image counts: - 1,430 images – `terrain_g05_mid_v1`- 10,000 images – `grid_g05_mid_v2`- 2,354 images – `terrain_g05_low_v1`- 3,707 images – `terrain_g05_high_v1`- 880 images – `terrain_g005_mid_v1`- 2,127 images – `terrain_g005_low_v1`- 11,325 images – `grid_g005_mid_v2`- 1,212 images – `terrain_g005_high_v1`- 348 images – `terrain_g1_mid_v1`- 4,285 images – `terrain_g1_low_v1`- 904 images – `terrain_g1_high_v1`- 3,000 images – `grid_g005_mid_v1`- 2,997 images – `grid_g005_low_v1`- 4,000 images – `grid_g005_high_v1`- 7,000 images – `grid_g05_mid_v1`- 7,098 images – `grid_g05_low_v1`- 7,000 images – `grid_g05_high_v1` Citation If you find SynRS3D useful in your research, please consider citing: @article{song2024synrs3d, title={SynRS3D: A Synthetic Dataset for Global 3D Semantic Understanding from Monocular Remote Sensing Imagery}, author={Song, Jian and Chen, Hongruixuan and Xuan, Weihao and Xia, Junshi and Yokoya, Naoto}, journal={arXiv preprint arXiv:2406.18151}, year={2024} } Contact For any questions or feedback, feel free to reach out via email: song@ms.k.u-tokyo.ac.jp. Enjoy using SynRS3D!