MIRAGE

Metric inference × Gaussian estimation

MIRAGE
Metric Inference and Reliability-Aware
Gaussian Estimation
from Feed-Forward Vision

TL;DR

MIRAGE teaser: Oxford Spires Gaussian reconstruction, trajectory, rendering and an indoor benchmark
On Oxford Spires Blenheim 02: lower ATE than COLMAP (3.27 vs. 5.50 cm), and higher rendering PSNR than Gaussian-LIC2 (27.82 vs. 21.28 dB). Scene-specific results, not dataset-wide averages.

01 / In motion

Every scene. A different perspective.

Browse Oxford Spires, M2DGR, INS, and our indoor benchmark. Each demo pairs a reconstruction-process illustration with a refined Gaussian rendering replay.

02 / Method

Grounded in motion.
Refined in appearance.

Feed-forward visual evidence meets LiDAR–inertial metric motion, with reliability-aware pose refinement and structure-aware Gaussian adaptation.

MIRAGE method pipeline from multimodal input to pose estimation and Gaussian reconstruction
The MIRAGE pipeline.

Quantitative results

Metric accuracy. Rendering fidelity.

Pose estimation ATE (cm) ↓

Rendering quality PSNR ↑ / SSIM ↑ / LPIPS ↓

03 / Beyond short trajectories

A long-horizon indoor benchmark.

Object-rich laboratories, dining areas, corridors, and offices. Synchronized multimodal observations and MIRAGE-established 6-DoF reference trajectories support evaluation of vision-only reconstruction.

10sequences
71,536RGB frames
3,576 srecorded motion
1.15 kmtrajectory length
Indoor benchmark camera images, LiDAR maps, trajectories and IMU signals
RGB observations, colored LiDAR maps, trajectories, and IMU measurements.
Explore six additional sequences +
Six additional indoor benchmark sequences

Benchmark evaluation

04 / Visual results

Qualitative comparisons.

Main-paper qualitative comparisons of pose trajectories and Gaussian rendering
Explore additional qualitative comparisons on Oxford Spires
Oxford Spires qualitative comparisons from the paper appendix
Explore additional qualitative comparisons on M2DGR & INS
M2DGR and INS qualitative comparisons from the paper appendix

Citation

BibTeX
@misc{wang2026mirage,
  title  = {{MIRAGE}: Metric Inference and Reliability-Aware Gaussian Estimation from Feed-Forward Vision},
  author = {Wang, Xinzhe and Jiang, Changjian and Song, Kaiwen and Li, Xudong and
            Ren, Kerui and Yi, Ran and Ma, Lizhuang and Shen, Chunhua and
            Xu, Linning and Lu, Tao and Yu, Mulin},
  year   = {2026},
  url    = {https://hiroxzwang.github.io/publications/mirage}
}