MvFFN: Multi-view Floor plan Feed-forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction
Reconstructing accurate camera poses and floor-plan layouts from sparse, unposed, wide-baseline RGB panoramas remains a challenging open problem. Direct multi-room floor-plan prediction is especially difficult in this setting: pose, geometry, semantics, and global layout structure are tightly…
This article presents MvFFN, a novel approach for reconstructing floor plans from unposed wide-baseline RGB panoramas. The method uses a multi-view feed-forward network to predict camera poses, geometry, and semantic outputs, followed by a post-processing pipeline to assemble multi-room layouts.
What happened
The article introduces MvFFN, a multi-view feed-forward network designed to reconstruct floor plans from unposed wide-baseline RGB panoramas. The approach uses local predictions from each panorama column, which are then aggregated to form global layouts. A post-processing pipeline assembles these predictions into multi-room floor-plans with consistent structures.
Why it matters
This work advances the field of computer vision by providing a more accurate method for floor-plan reconstruction, which is crucial for applications like robotics, augmented reality, and interior design. The use of local predictions and aggregation improves the accuracy of global layouts, offering a significant improvement over existing methods.
- MvFFN's local prediction approach leads to more accurate global floor-plan reconstructions.
- The use of a post-processing pipeline ensures consistent and connected room structures in the final layouts.
- The ZInD-CrossView dataset offers a new benchmark for research in multi-view floor-plan reconstruction.
The bigger picture
The proposed method improves the accuracy of floor-plan reconstruction, which can enhance applications requiring precise spatial understanding, such as robotics and augmented reality. The introduction of the ZInD-CrossView dataset provides a valuable resource for future research in this area.
MvFFN presents an effective solution for reconstructing multi-room floor-plans from unposed wide-baseline RGB panoramas. By leveraging local predictions and a dedicated post-processing pipeline, the method achieves superior accuracy compared to existing approaches, with the added benefit of a new dataset for future research.
Technical details
- Multi-view Ordered Wall Instance Segmentation (MOIS): Each panorama column predicts its observed wall and the next two adjacent walls in clockwise order.
- Multi-view Floor-plan Feed-Forward Network (MvFFN): Jointly predicts coarse camera poses, dense global geometry, and per-wall semantic outputs.
- Layout Chaining (LC) post-processing pipeline: Assembles predictions into multi-room floor-plans with shared global coordinates and consistent wall identities.
- ZInD-CrossView dataset: Augments ZInD with globally unique wall instances and cross-view segmentation labels for the task.
FAQ
What is MvFFN?
MvFFN is a Multi-view Floor-plan Feed-Forward Network designed to reconstruct accurate camera poses and floor-plan layouts from unposed wide-baseline RGB panoramas.
How does MvFFN work?
MvFFN uses Multi-view Ordered Wall Instance Segmentation (MOIS) to make local predictions for each panorama column, which are then aggregated to form global layouts. A post-processing pipeline called Layout Chaining (LC) assembles these predictions into multi-room floor-plans.
What is the ZInD-CrossView dataset?
The ZInD-CrossView dataset is an augmented version of ZInD that includes globally unique wall instances and cross-view segmentation labels, specifically designed for multi-view floor-plan reconstruction tasks.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv - cs.CV (CV) on Oct 14, 2026.
Comments
Sign in to join the discussion.