Research map / 3D vision / 3D generation
BADGR: Bundle Adjustment Diffusion Conditioned by Gradients for Wide-Baseline Floor Plan Reconstruction
Read the paperCheck your idea against 108,967 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.
Proposes BADGR, a conditional diffusion model that takes dense per-column outputs of a single-step Levenberg-Marquardt planar bundle adjustment layer to jointly refine camera poses and floor plan layouts from sparse panoramas.
Problem
Accurately recovering camera poses and floor plan layouts from sparsely captured, wide-baseline 360° panoramas, where co-visibility is limited and coarse initial poses/layouts are view-inconsistent.
Approach
Instead of guided diffusion (gradients applied only at inference), the model is conditioned during training on dense per-column bundle adjustment adjustments from a differentiable planar BA layer, and trained with a reprojection loss plus layout denoising, using a transformer denoiser with angle-constrained wall representation.
Key results
- On ZInD with 0.6 images per room, BADGR reduces mean pose translation error from 22.9 cm (CovisPose) / 20.7 cm (CovisPose+) to 12.2 cm and mean visible wall error from 14.0 cm to 7.1 cm.
- Conditional BADGR outperforms the guided-diffusion baseline BA+DM at 1 image per room (visible wall mean 7.0 cm vs 8.6 cm; pose 11.2 cm vs 15.0 cm).
- On RPLAN with 2% max boundary noise, BADGR achieves 0.59% mean pose error versus 1.47% for BA-Only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Refine coarse poses and layouts with BADGR instead of CovisPose+ | ZInD reconstruction · ZInD test set, 0.6 images/room | Pose translation mean (cm): 20.7 -> 12.2 | Helped |
| Refine coarse poses and layouts with BADGR instead of BA-Only | ZInD reconstruction · ZInD test set, 0.6 images/room | Visible walls mean error (cm): 12.8 -> 7.1 | Helped |
| Use BADGR instead of BA-Only | RPLAN reconstruction · RPLAN test set, 1 image/room | Pose error mean (%): 0.58 -> 0.35 | Helped |
| Use BADGR instead of BA-Only | RPLAN reconstruction · RPLAN test set, 1 image/room | Visible layout error mean (%): 0.62 -> 0.33 | Helped |
| Add reprojection loss to BADGR with planar BA inputs (ablation) | BADGR · ZInD test set, 1 image/room | Visible walls mean error (cm): 7.2 -> 7.0 | Helped |
| Add reprojection loss to BADGR with planar BA inputs (ablation) | BADGR · ZInD test set, 1 image/room | Pose mean error (cm): 12.0 -> 11.2 | Helped |
| Add planar BA inputs while retaining reprojection loss (ablation) | BADGR · ZInD test set, 1 image/room | Visible walls mean error (cm): 8.4 -> 7.0 | Helped |
| Add planar BA inputs while retaining reprojection loss (ablation) | BADGR · ZInD test set, 1 image/room | Pose mean error (cm): 23.6 -> 11.2 | Helped |
| Use guided diffusion with BA adjustment added to diffusion adjustment instead of BADGR conditioning (ablation) | BA + DM · ZInD test set, 1 image/room | Pose mean error (cm): 15.0 -> 11.2 | Hurt |
Surprising findings
- BA-Only gives no benefit or worse results at low input density (below one image per room), while BADGR still improves over the starting point.
Limitations
- Requires pre-assigned column-to-wall correspondences and known wall count/connectivity, plus a human-in-the-loop step for global wall connectivity.
- FloorPlan-60K dataset is not publicly released; release pending owner (Zillow) permission.
- Assumes Atlanta World layouts with no curved walls and known relative camera heights.
Experimental setup
Datasets: FloorPlan-60K ZInD RPLAN
Benchmarks: ZInD RPLAN
Metrics: pose rotation error (deg) pose translation error (cm) visible wall error (cm) normalized distance error (%)
Baselines: CovisPose CovisPose+ BA-Only BA+DM (guided diffusion) Diffusion model without BA conditioning
Builds on: DDPM DDIM HouseDiffusion CovisPose Theseus
Closest papers by meaning
- MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction (NeurIPS 2026 · similarity 0.88)
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.85)
- CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage (NeurIPS 2026 · similarity 0.85)
- Seg2Reg: Differentiable 2D Segmentation to 1D Regression Rendering for 360 Room Layout Reconstruction (CVPR 2024 · similarity 0.84)
- Coherent 3D Scene Diffusion From a Single RGB Image (NeurIPS 2024 · similarity 0.84)
- DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusion (CVPR 2025 · 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.