Research map / 3D vision / Depth and stereo

MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Cost-volume multi-view stereo

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 MOIS, a local cross-view wall-instance formulation, and MvFFN, a two-branch two-stage feed-forward multi-view transformer, plus Layout Chaining post-processing to reconstruct multi-room floor plans and poses from unposed wide-baseline panoramas.

Problem

Recovering camera poses and multi-room floor-plan layouts from sparse, unposed, wide-baseline 360° panoramas, where direct global floor-plan prediction is hard to learn and error-prone.

Approach

Replace direct global wall-sequence prediction with dense anchor-conditioned local cross-view wall instances and within-room ordered wall connectivity (MOIS), predicted by a feed-forward multi-view transformer and then assembled into global layouts by deterministic Layout Chaining.

Key results

• On ZInD-XV 5-view, MvFFN reaches 88.1 wall F1 and 73.3 junction F1, versus 73.4 wall F1 for the best modular baseline (LGT-PP with CovisPose).
• Pose: MvFFN lowers mean 5-view position error to 12.5 cm vs 12.6 cm for CovisPose+GST and gives lower variance, though CovisPose+GST has lower median error.
• Planar bundle adjustment improves 68.0% of 5-view samples and 66.0% of 8-view samples on ZInD-XV, with mean matched-wall distance 14.8 to 12.3 cm on the improved subset.

Surprising findings

• MvFFN outperforms the LGT-PP baseline even when that baseline is given ground-truth poses, so gains are not solely from better pose estimation.
• Training with longer MOIS prediction windows (5 or 7 walls) consistently degrades layout accuracy.

Limitations

• Pose median error is worse than CovisPose+GST in the main ZInD-XV evaluation.
• BADGR-based refinement gives mixed results and is sensitive to initialization; overall refinement gains are modest.
• Relies on CovisPose-derived rotation and correspondence priors and on known relative camera heights.

Experimental setup

Datasets: ZInD-XV HM3D-XV ZInD HM3DSem

Benchmarks: ZInD-XV HM3D-XV

Metrics: wall precision wall recall wall F1 junction precision junction recall junction F1 matched-wall distance camera position error (cm) camera rotation error (deg) Pooled Ordered Wall-Slot IoU

Baselines: CovisPose+GST VGGT LGT-Net+CovisPose post-processing RoomFormer PolyRoom

Builds on: VGGT CovisPose BADGR LGT-Net DUSt3R

Closest papers by meaning

• BADGR: Bundle Adjustment Diffusion Conditioned by Gradients for Wide-Baseline Floor Plan Reconstruction (CVPR 2025 · similarity 0.88)
• HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.87)
• CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.86)
• CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage (NeurIPS 2026 · similarity 0.86)
• MultiDiff: Consistent Novel View Synthesis from a Single Image (CVPR 2024 · similarity 0.85)
• Unposed Sparse Views Room Layout Reconstruction in the Age of Pretrain Model (ICLR 2025 · similarity 0.85)