# BADGR: Bundle Adjustment Diffusion Conditioned by GRadients for Wide-Baseline Floor Plan Reconstruction

**Authors:** Yuguang Li, Ivaylo Boyadzhiev, Zixuan Liu, Linda Shapiro, Alex Colburn  
**arXiv:** [2503.19340](https://arxiv.org/abs/2503.19340)  
**Date:** 2025-03-25  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2503.19340v1/

## Key points

- BADGR reconstructs camera poses and floor plans from RGB panoramas captured at widely separated viewpoints.
- The method jointly refines camera poses and room layouts using a gradient-conditioned diffusion model, starting from a coarse reconstruction and 1D floor-boundary predictions.
- A single-step Levenberg–Marquardt optimizer generates dense per-entity outputs that condition diffusion and help reduce reprojection errors across views.
- BADGR complements bundle adjustment: learned structural priors, such as wall adjacency and collinearity, help stabilize pose estimation when cross-view features are sparse or unreliable.
- Global context is intended to limit the effect of errors in local boundary predictions.
- The model is trained exclusively on 2D floor plans; the authors say this simplifies data acquisition, enables robust augmentation, and supports varying input densities.
- The authors report that BADGR significantly outperforms prior methods across experiments with different numbers of input images, but the abstract provides no metrics, scores, datasets, or baseline details.
- The abstract does not specify failure cases, capture or geometry assumptions, generalization, computational cost, or sensitivity to boundary predictions; assessing these requires the full paper.

## Summary

BADGR addresses the challenge of reconstructing camera poses and floor-plan layouts from RGB panoramas captured from widely separated viewpoints. Large viewpoint changes make it difficult to match observations across images, and errors in camera position or orientation can distort the inferred room geometry. The method combines image-based evidence with learned priors about plausible floor-plan structures. It jointly refines camera poses and layouts using a gradient-conditioned diffusion model. Starting from a coarse reconstruction, it uses one-dimensional floor-boundary predictions from a variable number of panoramas. A single-step Levenberg–Marquardt optimizer produces dense, per-entity outputs that condition the diffusion process. BADGR predicts camera and wall positions while reducing reprojection errors, encouraging consistency between the reconstructed scene and observations across views. The authors present diffusion-based layout generation as complementary to bundle adjustment. Bundle adjustment uses features visible in multiple images, but those features may be sparse or unreliable when viewpoints are far apart. BADGR’s learned structural priors—including wall adjacency and collinearity—provide additional constraints that can help infer spatial relationships and stabilize the pose graph. The model’s global context is also intended to reduce the influence of errors in local boundary predictions. BADGR is trained exclusively on 2D floor plans. The authors say this simplifies data acquisition, supports robust augmentation, and accommodates different input densities. They report experiments with varying numbers of input images and state that the method significantly outperforms the state of the art in pose and floor-plan reconstruction. The abstract provides no dataset names, baseline results, metric definitions, numerical scores, or measurements of the improvements. It therefore does not support a detailed account of the experimental results or key figures. It also does not specify failure cases, assumptions about panorama capture or room geometry, generalization to unfamiliar layouts, computational cost, or sensitivity to boundary predictions. Assessing these aspects requires the full paper, which is unavailable here.