baseline_f1 float64 | idea3 dict | total_time float64 | idea6 dict | idea3p dict | idea1 dict | idea2 dict | idea4 dict | idea5 dict |
|---|---|---|---|---|---|---|---|---|
0.662039 | {
"encoder_path": "saved_model_files/enc_synth_i3.pt",
"best_test_f1": 0.810963271319175,
"best_loss": 2.927182441533998
} | 311.3 | {
"encoder_path": "saved_model_files/enc_synth_i6_k5.pt",
"best_test_f1": 0.779650552076869,
"k": 5,
"delta": 0.117612
} | {
"encoder_path": "saved_model_files/enc_synth_i3p_t200.pt",
"best_test_f1": 0.8081918680510749,
"t0": 200,
"delta": 0.146153
} | {
"encoder_path": "saved_model_files/enc_synth_i1.pt",
"best_test_f1": 0.7727300740910227,
"method": "conditional_ddpm+interpolation",
"delta": 0.110691
} | {
"encoder_path": "saved_model_files/enc_synth_i2.pt",
"best_test_f1": 0.6683493661541913,
"method": "vae+style_sampling",
"delta": 0.006311
} | {
"encoder_path": "saved_model_files/enc_synth_i4_t200_k0.5.pt",
"best_test_f1": 0.7379752629133297,
"n_kept": 5436,
"t0": 200,
"keep_frac": 0.5,
"delta": 0.075936
} | {
"encoder_path": "saved_model_files/enc_synth_i5.pt",
"best_test_f1": 0.7444511820420756,
"best_round": 1,
"rounds": 3,
"delta": 0.082412
} |
STER: Spatial-aware Three-dimensional Entity Resolution
STER extends the 3dSAGER benchmark (SIGMOD 2026, Geospatial Entity Resolution over 3D Objects) towards an AAAI submission. We build, on top of the exact same The Hague ER benchmark, a suite of 8 method proposals spanning four scenarios × two modern method families:
| Scenario | Diffusion-based | World-Model-based |
|---|---|---|
| Long-tail | 1A Conditional Diffusion p(B|A) |
1B Disentangled 3D Geometric World Model |
| Noisy | 2A Geometric Grammar Correction | 2B Causal World Model (signal⊥noise) |
| Few-shot | 3A Transformation-aware Bridge | 3B JEPA Cross-source Geometric Imagination |
| Zero-shot | 4A Denoising Trajectory Signature | 4B Flow Matching on the Building Manifold |
Core design principle: model everything in the 25-dim geometric property space (not raw 3D mesh space), so diffusion / flow / world-model training drops from days to minutes and stays fully comparable with the original 3dSAGER pipeline.
Ground truth without manual labels
Both data sources (The Hague 3D City Model = candidates; 3DBAG = index) carry BAG building identifiers. Same BAG id ⇒ true match. This yields large-scale ground truth at zero labeling cost.
Repository layout
STER/
├── README.md # this card + live progress board
├── SPEC/ # data acquisition & build specification (feasibility-checked)
├── code/ # crawlers, property extraction, variant builders, 8 methods
├── data/ # derived property vectors + scenario splits (raw stays at source)
├── experiments/ # configs, result tables, ablations
└── logs/ # progress log & milestone records
Data sources (all verified reachable)
- The Hague 3D City Model 2022 — municipality open data (candidate set
D_C). - 3DBAG —
api.3dbag.nl, OGC API Features, per-building LoD1.2/1.3/2.2 (index setD_I). - Official 3dSAGER code + released partitions —
github.com/BarGenossar/3dSAGER.
We redistribute only derived features / splits + acquisition scripts + checksums; original bulk data is fetched from the sources under their licenses (3DBAG: CC BY 4.0).
Status
See logs/progress.md for the live milestone board. This repository is updated continuously as data,
records, experiment results, code, and deployment artifacts are produced.
- Downloads last month
- 112