STER: Zero-shot 3D Geometric Entity Resolution Benchmark

Multi-city, cross-LoD 3D building matching benchmark for the NS-D2S paper (AAAI 2026). Strictly follows the 3dSAGER (SIGMOD 2026) methodology and data format.

Dataset Overview

Dataset	City	Country	Buildings	LOD Source	Urban Typology
amsterdam	Amsterdam	NL	123,259	3DBAG LOD1.2/1.3/2.2	Historic canal city
rotterdam	Rotterdam	NL	152,694	3DBAG LOD1.2/1.3/2.2	Post-war modern
hague	Den Haag	NL	181,491	3DBAG LOD1.2/1.3/2.2	Administrative center
utrecht	Utrecht	NL	121,293	3DBAG LOD1.2/1.3/2.2	Historic university city
eindhoven	Eindhoven	NL	181,811	3DBAG LOD1.2/1.3/2.2	Modern industrial
groningen	Groningen	NL	72,696	3DBAG LOD1.2/1.3/2.2	Northern university town
maastricht	Maastricht	NL	73,147	3DBAG LOD1.2/1.3/2.2	Historic small city
chiyoda	Chiyoda, Tokyo	JP	8,438	PLATEAU LOD1/LOD2	Government/business core
chuo	Chuo, Tokyo	JP	8,507	PLATEAU LOD1/LOD2	Historic commercial (Ginza)
shinjuku	Shinjuku, Tokyo	JP	2,947	PLATEAU LOD1/LOD2	Skyscraper/entertainment
setagaya	Setagaya, Tokyo	JP	8,415	PLATEAU LOD1/LOD2	Residential suburban
minato	Minato, Tokyo	JP	9,023	PLATEAU LOD1/LOD2	Business/embassy district
bunkyo	Bunkyo, Tokyo	JP	3,334	PLATEAU LOD1/LOD2	Academic/residential
koto	Koto, Tokyo	JP	5,279	PLATEAU LOD1/LOD2	Waterfront/new development
ota	Ota, Tokyo	JP	1,318	PLATEAU LOD1/LOD2	Mixed industrial/residential
kyoto	Kyoto	JP	34,831	PLATEAU LOD1/LOD2	Historic ancient capital
osaka	Osaka	JP	3,945	PLATEAU LOD1/LOD2	Metropolitan commercial
sakai	Sakai	JP	4,128	PLATEAU LOD1/LOD2	Satellite industrial city

Total: 18 cities, 996,590 buildings

Data Format

Each city directory (data/{city}/) contains exactly 5 files:

File	Description
object_dict_raw.joblib	Primary data: SIGMOD-compatible 3D building meshes
manifest.json	Per-city metadata (bbox, filter params, timestamps)
{city}_seed1.pkl	Train/test partition (seed 1)
{city}_seed2.pkl	Train/test partition (seed 2)
{city}_seed3.pkl	Train/test partition (seed 3)

object_dict_raw.joblib (SIGMOD Format)

{
    'cands': {            # LOD1/LOD1.2 (coarse source)
        building_id: {
            'polygon_mesh': [[[x,y,z],...], ...],
            'vertices': np.array([[x,y,z], ...]),   # (N, 3)
            'centroid': np.array([x, y, z]),
        }, ...
    },
    'index': {            # LOD2/LOD2.2 (detailed source)
        building_id: { ... }, ...
    },
    'mapping_dict': {     # integer index ↔ building ID
        'cands': {0: id1, 1: id2, ...},
        'index': {0: id1, 1: id2, ...},
    },
    'inv_mapping_dict': { # building ID ↔ integer index
        'cands': {id1: 0, id2: 1, ...},
        'index': {id1: 0, id2: 1, ...},
    },
}

Partition Files ({city}_seed{1,2,3}.pkl)

{
    'train': {
        'negative_sampling': {
            'small': {2: [(cand_idx, index_idx), ...], 5: [...]},
            'large': {2: [...], 5: [...]}
        }
    },
    'test': {
        'matching': {
            'negative_sampling': {'small': {...}, 'large': {...}},
            'blocking-based': {'small': {2: [...], 5: [...]}, 'large': {...}}
        },
        'blocking': {
            'small': {'cands': {idx,...}, 'index': {idx,...}},
            'large': {'cands': {idx,...}, 'index': {idx,...}}
        }
    }
}

80/20 spatial grid-based train/test split. 3 random seeds for statistical significance.

Quick Start

import joblib

# Load dataset
od = joblib.load('data/amsterdam/object_dict_raw.joblib')

# Compute 25 geometric properties via 3dSAGER pipeline
from object_properties import ObjectPropertiesProcessor
proc = ObjectPropertiesProcessor(od, vector_normalization=True)
# proc.prop_vals_dict is ready for PairProcessor → classifier training

# Load partition
with open('data/amsterdam/amsterdam_seed1.pkl', 'rb') as f:
    partition = pickle.load(f)
train_pairs = partition['train']['negative_sampling']['large'][2]

Cross-LoD Matching Paradigm

• Dutch cities: LOD1.2 (cands) vs LOD2.2 (index), shared BAG ID = GT match
• Japan cities: LOD1 (cands) vs LOD2 (index), shared PLATEAU building ID = GT match

Same building ID across LODs = positive match. No human annotation needed.

25 Geometric Properties

bounding_box_width, bounding_box_length, area, perimeter, perimeter_ind, volume, convex_hull_area, convex_hull_volume, ave_centroid_distance, height_diff, num_floors, axes_symmetry, compactness_2d, compactness_3d, density, elongation, shape_ind, hemisphericality, fractality, cubeness, circumference, aligned_bounding_box_width/length/height, num_vertices

Log-normalized via official ObjectPropertiesProcessor.

Filtering

• Buildings with <10 polygon faces excluded
• Dutch: combined LOD1.2+LOD1.3+LOD2.2 ≥10 faces
• Japan: LOD1 ≥10 and LOD2 ≥10 faces

Coordinate Systems

• Dutch: EPSG:28992 (Amersfoort / RD New)
• Japan: JGD2011 (EPSG:6697)

Code

Script	Description
code/object_properties.py	25 geometric property computation (3dSAGER official)
code/build_benchmark.py	3DBAG Dutch multi-city builder
code/build_plateau.py	PLATEAU Japan city builder
code/fix_sigmod_v2.py	Format unification (SIGMOD object_dict)
code/generate_partitions.py	Train/test partition generator
code/pipelines.py	3dSAGER pipeline (matching/blocking)
code/main.py	3dSAGER entry point
code/ster_*.py	NS-D2S experiment scripts

Citation

If you use this benchmark, please cite both this dataset and the original 3dSAGER paper.