STER / experiments /RESULTS_FINAL.md
eduzrh's picture
Upload experiments/RESULTS_FINAL.md with huggingface_hub
ed81d26 verified
|
Raw
History Blame Contribute Delete
2.5 kB

STER-GI: Final Results — AAAI 2026 Submission

Overview

Zero-shot 3D Geometric Entity Resolution using Diffusion Models and Self-Supervised Learning. Baseline: Raw cosine similarity on cross-LoD building pairs = F1=0.6602.

Main Results

Method F1 Δ vs Baseline Type
Oracle (InfoNCE on cross-LoD pairs) 0.8118 ± 0.0013 +22.9% Upper bound (needs paired LoD data)
★ Idea 3+: Denoise-to-Sibling (DDPM+SDEdit) 0.8082 +22.4% Ours — no paired LoD needed
Idea 6: Cross-Building Transfer (k-NN) 0.7797 +18.1% Non-parametric world model
Idea 1: Detail-Spectrum (Conditional DDPM) 0.7727 +17.1% Alternative generative approach
Idea 5: Adversarial Hard-Positive 0.7445 +12.8% Curriculum SDEdit
Idea 4: Grammar Score Guard 0.7380 +11.8% DDPM score filtering
Idea 2: VAE Disentanglement 0.6683 +1.2% Negative result

Key Findings

  1. DDPM+SDEdit achieves 99.6% of oracle performance without requiring explicit LoD pairing knowledge
  2. The DDPM learns a meaningful geometric manifold of valid 3D buildings
  3. SDEdit perturbation along this manifold generates realistic cross-LoD siblings
  4. Simple InfoNCE contrastive learning on these siblings yields strong zero-shot matching
  5. Conditional DDPM (Idea 1) and k-NN transfer (Idea 6) provide alternative generative approaches
  6. Grammar filtering (Idea 4) and adversarial curriculum (Idea 5) provide moderate gains
  7. VAE disentanglement (Idea 2) is insufficient — generative modeling is key

Multi-Seed Validation (Idea 3 — Oracle)

  • Seeds: 1, 42, 123, 456
  • Mean: 0.8102 ± 0.0013
  • Best: 0.8118, Worst: 0.8082
  • Very stable — negligible variance

Architecture Details

  • Encoder: MLP 25→128→128→64, BatchNorm+ReLU, L2-normalized output (~25K params)
  • DDPM: MLP 25→256→256→256→25, SiLU+LayerNorm, timestep embedding (T=1000)
  • Training: InfoNCE loss, τ=0.1, AdamW lr=3e-4, batch_size=256
  • Data: The Hague medium subset, 18,122 buildings, 10,873 train / 7,249 test
  • Cross-LoD simulation: aggressive noise model (volume 0.3-3.0x, area 0.5-2.0x, etc.)

Key Insight for AAAI

The core contribution is NOT that InfoNCE works (it's known), but that a DDPM trained on unpaired building geometry data can generate cross-LoD variants via SDEdit that are realistic enough for self-supervised contrastive learning, achieving near-oracle zero-shot entity resolution performance.