Research map / 3D vision / 3D generation
Towards Reconstructing Geographically Diverse Architecture with 3D Foundation Models
Nicky Kriplani, Yiwen Zhang, Angelina Wang, Hadar Averbuch-Elor
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ArchWorld, a geographically diverse benchmark of 3D architectural landmarks, and uses it to analyze errors of current 3D foundation models in single feed-forward scene reconstruction.
Problem
Internet-scale photo collections of landmarks are highly challenging for 3D foundation models that estimate scene structure in a single feed-forward pass, and their failure modes on geographically diverse architecture are not well characterized.
Approach
A benchmark of 3D architectural landmarks with high geographic coverage and rich metadata, used for a thorough error analysis of current 3DFMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply ArchAttend (gradient-ascent step on token representations toward anchor architectural element) | DA3 · ArchWorld, all scenes | median normalized Chamfer distance: 10.36 -> 8.92 | Improved |
| apply ArchAttend | DA3 · ArchWorld North America | median normalized Chamfer distance: 18.83 -> 12.82 | Improved |
| apply ArchAttend | DA3 · ArchWorld Tower architecture | median normalized Chamfer distance: 21.04 -> 16.29 | Improved |
| apply ArchAttend | DA3 · ArchWorld regional disparity | disparity (max minus min regional median): 11.26 -> 6.96 | Improved |
| apply ArchAttend | DA3 · ArchWorld architecture disparity | disparity (max minus min type median): 13.75 -> 9.14 | Improved |
| use a single scene-agnostic step size eta instead of best of five (ablation) | ArchAttend on DA3 · ArchWorld | regional gap closed (%): closes about 41.7% of the regional gap | Improved |
| train on more data from a region (data representation) | 3DFMs (VGGT, DA3, pi3) · ArchWorld regional performance | regional error vs training representation | No clear effect |
| normalize Chamfer distance by scene bounding box size (ablation) | 3DFM evaluation · ArchWorld regional disparities | correlation with data representation | Mixed |
| use stronger overall 3DFM (pi3) | 3DFMs · ArchWorld regions | geographic disparity: pi3 East Asia 2.75 vs Southeast Asia 1.34 | No clear effect |
| add construction year as predictor of error | 3DFMs · ArchWorld (149 annotated scenes) | reconstruction error | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Experimental setup
Benchmarks: ArchWorld
Closest papers by meaning
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.86)
- SpatialBench: Is Your Spatial Foundation Model an All-Round Player (NeurIPS 2026 · similarity 0.86)
- Unified Panoramic Geometry Estimation via Multi-View Foundation Models (NeurIPS 2026 · similarity 0.86)
- Feat2GS: Probing Visual Foundation Models with Gaussian Splatting (CVPR 2025 · similarity 0.85)
- Scalable Adaptation of 3D Geometric Foundation Models via Weak Supervision from Internet Video (NeurIPS 2026 · similarity 0.85)
- Large Spatial Model: End-to-end Unposed Images to Semantic 3D (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
