GeoRel-100k

GeoRel-100k is a large-scale synthetic dataset of 3D object arrangements rendered in Blender, designed for research on geometric and spatial relations between objects. Each arrangement provides multi-view photo-realistic renderings together with full geometric ground truth (depth, surface normals, per-object segmentation), camera calibration in COLMAP convention, the original Blender source file, and a structured scene graph describing pairwise spatial relations and metric distances between every pair of objects in the scene.

Dataset structure

The dataset is shipped as one ZIP archive per scene under part1/ as samples and 'full/' as the complete dataset:

part1/
  scene_001.zip
  scene_002.zip
  ...
  scene_087.zip

Each scene_NNN.zip contains one or more independent arrangements arr_MMM/ of the same set of objects. A single arrangement has the following layout:

arr_001/
  scene.blend                 # Blender source
  scene_graph.json            # objects, pairwise relations, distances, adjacency
  metadata.json               # render config, camera intrinsics + extrinsics, object bboxes
  rgb/                        # camNN_az<az>_el<el>.png (PNG, 512x512)
  depth/                      # *_zdepth.exr  and  *_euclidean.exr (OpenEXR)
  normals/                    # *_worldspace.exr  and  *_camspace.exr (OpenEXR)
  masks/                      # *_objNNN.exr per-object + *_combined.exr id-map (OpenEXR)
  colmap/                     # cameras.txt, images.txt, points3D.txt

Filenames for renders encode camera id, azimuth and elevation, e.g. cam03_az128.2_el32.0.png.

scene_graph.json

Field	Description
scene_id	scene_NNN_arr_MMM
num_objects	Number of objects in the arrangement
nodes[]	id, object_name, class, center_3d_m, bbox_min_m, bbox_max_m, dimensions_m, footprint_radius_m, material_name
edges[]	source, target, distances_m {center_3d, center_2d, surface_to_surface, delta_z}, relations {proximity, vertical, horizontal}
adjacency	Symmetric {object_id: {neighbor_id: surface_to_surface_distance_m}}

metadata.json

Field	Description
render	resolution_x, resolution_y, num_cameras
cameras[]	Per-camera intrinsics (fx_px, fy_px, cx_px, cy_px, fov_h_deg, fov_v_deg) and extrinsics (position_world_m, R_w2c_colmap, t_w2c_colmap_m) in COLMAP world-to-camera convention
objects[]	Per-object 3D bounding boxes

Usage

Download a single scene with huggingface_hub:

from huggingface_hub import hf_hub_download

zip_path = hf_hub_download(
    repo_id="georel-100k/GeoRel-100k",
    repo_type="dataset",
    filename="part1/scene_001.zip",
)

Or stream the whole dataset with datasets:

from datasets import load_dataset
ds = load_dataset("georel-100k/GeoRel-100k", streaming=True)

A machine-readable dataset description following the MLCommons Croissant 1.0 specification is provided at croissant.json.

License

Released under CC BY 4.0.

Citation

@misc{georel100k_2026,
  title  = {GeoRel-100k: A Synthetic Dataset for Geometric Object Relations},
  author = {Anonymous},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/georel-100k/GeoRel-100k}}
}