CHI25 — Sketch2Terrain: AI-Driven Real-Time Terrain Sketch Mapping in Augmented Reality — Dataset

The dataset for CHI2025 paper: Sketch2Terrain: AI-Driven Real-Time Terrain Sketch Mapping in AR.

In this paper, we present a generative AI model (Pix2Pix model) that converts 3D sketches created in AR into 3D terrain models. The GitHub page for this open-source project is https://github.com/TX-XR/Sketch2Terrain.

Introduction: The dataset contains the pairs of human-sketched main structures of the terrain and the corresponding height map of the ground truth terrain. The dataset uses real terrain features from all over the world, including rivers, ridges, basins, and peaks, as training data.

Data Collection: We collected 287 sketches, created by two researchers at a rate of 10 sketches per hour using the AR 3D sketching tools we developed. Both sketch and terrain height map values are normalized to a range of [0, 255] and saved as PNG files. 

Data Augmentation: The collected sketch maps were split into training (80%) and testing (20%) subsets, with augmentation applied only to the training set. Augmentation included height value scaling and tailored transformations like translation and rotation, while blank spaces created during transformations were filled with black pixels. Augmentation increased the training dataset to over 2,000 samples.

Feel free to distribute and use the data.