cs.CVOct 26, 2026

Metric Scale for Insta360 Structure-from-Motion from the Camera's Embedded IMU

Authors: Davide Cuttini, Diego Di Tommaso, Alessio Antoniacomi, Luca Simonetti

Organizations: Over The Reality

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Structure-from-motion (SfM) from the dual-fisheye video of a consumer 360-degree camera recovers geometry only up to a similarity transform; in the pipeline considered here, metric scale is obtained from smartphone ARKit or ARCore poses captured alongside the camera. This paper evaluates a self-contained alternative: the IMU embedded in the Insta360 X5 and X6, recorded at 1 kHz. The sensor is characterised (noise density, camera-IMU synchronisation, accelerometer scale factor, bias stability), the camera-IMU rotation and frame timing are estimated from the data, and the scale is recovered by a linear visual-inertial alignment over the SfM poses of a whole survey. The standard formulation, with gravity and accelerometer bias shared across the survey, overestimates the scale by 3.4% on average; estimating each contiguous segment independently and averaging removes the bias. On nine surveys with a co-mounted LiDAR scanner the inertial scale deviates from the LiDAR by 0.74% in mean absolute error, against 1.94% for the smartphone scale; on printed fiducial markers both agree within the 1% uncertainty of the target. On the reference trajectories of thirty sequences of a public benchmark, brought to scales of 1/40, 1 and 40, the estimator recovers the metric scale with a mean error of -0.07% and a standard deviation of 1.28%. On fifty production surveys where the smartphone is the only reference, the two scales agree to a standard deviation of 2.6%, the inertial scale being 0.9% lower on average; the difference does not depend on window construction, grows with the extent of the smartphone anchor, and cannot be attributed to either reference. An ablation over window construction and a comparison with monocular metric-depth networks complete the study.

Figures & tables

setsurveysdurationpath lengthreference
LiDAR9100 to 228 s101 to 241 mco-mounted LiDAR; ARKit (1) and ARCore (8)
Fiducial1203 s159 mArUco markers; ARKit
Production A1253 to 600 s49 to 775 mARKit
Production B50246 to 708 s192 to 1254 m (33 km in total)ARKit
Public benchmark3069 to 276 s72 to 352 m (5.5 km in total)LiDAR-inertial trajectory
Table 1: Survey sets used in the evaluation. The models of the first four sets are dual-fisheye rig reconstructions from the production pipeline; for the public benchmark the input to the estimator is the reference trajectory.
Figure 1: Amplitude spectral density of the gyroscope y axis for the Insta360 X5 (blue) and X6 (orange) over 80 s of hand-held recording, estimated with Welch’s method (segment length 4096). The band below 50 Hz is dominated by operator motion; above 100 Hz only sensor noise remains. The X6 is flat (white noise); the X5 is attenuated by a firmware low-pass filter.
Figure 2: Left: angular rate about the dominant axis over 10 s of a hand-held sequence with fast rotations, from video-only rotation estimation (blue) and from the gyroscope after axis mapping (orange). Right: mean absolute correlation between the two as a function of the time shift applied to the gyroscope; the maximum lies at −- 2 ms, within the ±5 ms resolution of the test, showing that the IMU timestamps stored in the file are aligned to the video.
surveycampaignpathsmartphone errorIMU error
L1first112 m+2.92%−- 2.11%
L2second140 m−- 1.18%+2.06%
L3second105 m−- 4.36%+0.02%
L4second106 m+0.63%+1.27%
L5second101 m−- 2.79%+0.39%
L6second233 m+0.68%−- 0.10%
Table 2: Scale error relative to the co-mounted LiDAR on nine surveys, with path lengths in LiDAR metres, for the smartphone reference that anchors the reconstruction and for the inertial estimate. Mean absolute error: 1.94% (smartphone), 0.81% (IMU). With lens 0 instead of lens 1 the IMU mean absolute error is 0.76%.
Figure 3: LiDAR survey L1, with the co-mounted LiDAR as reference. Left: camera trajectory from the production SfM model, scaled by the IMU estimate and aligned to the recovered gravity direction, coloured by elapsed time (110 m at the IMU scale, 112 m in LiDAR metres, over 143 s of registered frames). Right: scale error relative to the LiDAR (solid line at zero) of the inertial estimate on each contiguous segment (bars), of the mean of the segments with its 95% confidence interval (dashed blue line and band, −- 2.11%) and of the ARKit reference (dashed orange, +2.92%), as in Table 2.
Figure 4: Scale error relative to the LiDAR reconstruction on the nine co-mounted surveys, for the inertial scale and for the smartphone reference that anchors the reconstruction. Each point is one survey; the dotted lines mark the mean absolute error of each source. The monocular depth networks of Section 5.9, evaluated on survey L1 only, fall outside the range of this plot.
Figure 5: (a) The two ArUco sheets as seen in the fisheye frames, one viewed obliquely through glass (ids 0 and 1, excluded) and one on a rigid signboard (ids 2 and 3). (b) The signboard sheet rectified through the homography defined by the eight detected corners, with the nominal 140 mm marker outlines (green) overlaid; the black squares fall slightly inside the nominal outlines, consistent with a print scale about 1% below nominal.
Figure 6: Ratio between the IMU-derived scale and the ARKit-derived scale for the twelve surveys of production set A (A1 to A12), sorted by value. Filled circles are full-trajectory estimates, shaded bars the ±1σ\pm 1\sigma bootstrap interval, small grey circles the estimates from the first and second halves of each trajectory; the grey band is the mean ± one standard deviation across surveys.
Figure 7: Production set B, fifty surveys without external reference. Left: ratio between the IMU-derived scale and the ARKit-derived scale on the fifty surveys, sorted by value, with the uncertainty on the mean of the per-run estimates as error bar, the ±2% band shaded and the set mean dashed. Right: the same ratio against the mean length of the contiguous runs it is estimated on; the runs span 22 to 105 intervals and the correlation is −- 0.15, so the window-length bias of Section 5.6 does not account for the difference.
window constructionsurveys coveredmean errormean absolute errorworst
whole survey as one system10+3.39%3.43%6.05%
contiguous runs10−- 0.05%0.85%2.11%
30 intervals, stride 1510−- 0.20%0.89%1.95%
fixed, 30 intervals10−- 0.32%0.90%2.23%
fixed, 50 intervals10−- 0.24%0.90%1.91%
fixed, 25 intervals10−- 0.44%0.95%1.52%
Table 3: Scale error against external ground truth on the ten surveys with LiDAR or fiducial reference, for each window construction. “Surveys covered” is the number of surveys on which the construction yields at least one valid window.
Figure 8: Mean absolute error against ground truth on the ten surveys with external reference, for each window construction of Table 3. Orange: the single joint system; blue: the contiguous-run construction adopted here. Every windowed variant improves on the joint system by a factor between 2.5 and 4, and all of them lie within 0.6 percentage points of each other.
surveysceneMoGe-2 / ARKitMoGe-3 / ARKitIMU / ARKit
A10indoor0.832 ( −- 16.8%)0.839 ( −- 16.1%)0.955
A12indoor0.956 ( −- 4.4%)0.889 ( −- 11.1%)0.947
A1indoor0.993 ( −- 0.7%)0.974 ( −- 2.6%)1.035
A3indoor0.999 ( −- 0.1%)1.004 (+0.4%)1.021
A11outdoor0.773 ( −- 22.7%)0.924 ( −- 7.6%)0.953
A4outdoor0.834 ( −- 16.6%)1.178 (+17.8%)1.013
Table 4: Scale from monocular metric depth relative to the ARKit-scaled production model on ten surveys, 30 frames each, with the inertial ratio for comparison. The inertial ratio is the segment-wise estimate of Sections 5.2 and 5.3 for A7 and A5, and the joint estimate of Section 5.4 for the other surveys, whose models are no longer available.
Figure 9: Scale bias of MoGe-2 (circles) and MoGe-3 (squares) metric depth relative to the ARKit-scaled production model on ten surveys of set A, four indoor (blue) and six outdoor (orange), from 30 frames each; grey diamonds mark the inertial scale. Outdoor surveys are worse on average, but one indoor survey fails with both networks and the two networks disagree by up to 34% on the same scene.
More in this topic

Structure from Motion

Structure from Motion estimates camera poses and 3D scene geometry from overlapping images by matching visual features across views.

See all 46 papers

Explore similar work

CardsList
Sep 3, 2026

VI3: Grounding Pretrained 3D Foundation Models with Inertial Cues

3D foundation models (3DFMs) excel at predicting camera poses and dense depth from multiple views of a scene, showcasing strong zero-shot generalization. However, as metric scale is not observable from monocular images, their absolute scale predictions are typically inaccurate. Inertial measurement units (IMUs), present in most devices, naturally complement monocular cameras by observing scaled motion. We introduce VI3, a model-agnostic framework that metrically anchors a pretrained 3DFM using only IMU readings. VI3 initializes and preintegrates the IMU to obtain a metric motion reference, which is then used to recover the scale of the 3DFM outputs. Our method includes adaptable anchoring strategies tailored to diverse 3DFM architectures. Experiments on synthetic and real aerial datasets demonstrate that VI3 recovers metric scale without ground-truth supervision while preserving geometric consistency, acting as a fine refinement under well-conditioned motion and as a strong prior when motion is less informative.

Ernesto Lozano, Alberto Jaenal, Javier Civera

I3A, Universidad de Zaragoza, Spain

3D Foundation Models
AskPDF
Mar 22, 2026

Quantifying Motion Excitation for Metric Scale Observability in Monocular Visual-Inertial Odometry

Monocular visual-inertial odometry (VIO) cannot recover metric scale from vision alone; scale must be resolved through inertial measurements. We present a trajectory-dependent observability analysis showing that translational acceleration, produced by curvature, not constant-speed straight-line travel, is the fundamental source that couples scale to the inertial state. This relationship is formalized through the gravity-acceleration asymmetry in the IMU model, from which we derive rank conditions on the observability matrix and propose a lightweight excitation metric computable from raw IMU data. Controlled experiments on a differential-drive robot with a monocular camera and consumer-grade IMU validate the theory, with straight-line motion yielding 9.2% scale error, circular motion 6.4%, and figure-eight motion 4.8%, with excitation spanning four orders of magnitude. These results establish trajectory design as a practical mechanism for improving metric scale recovery.

Hadush Hailu, Bruk Gebregziabher, Siddhartha Gudipudi +1

Computer Science, Maharishi International University, Iowa, USA · Electrical Engineering and Computing, University of Zagreb, Zagreb, Croatia · Computer Science, Iowa State University, Iowa, USA +1

Inertial OdometryVisual-Inertial OdometryInertial Navigation
AskPDF
Oct 8, 2026

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

Feed-forward 3D priors enable high-quality reconstruction from image streams but often lack reliable metric scale and suffer from long-term trajectory drift. In contrast, LiDAR--inertial estimation provides metrically accurate motion but lacks the image-space consistency needed for high-fidelity reconstruction. We propose MIRAGEMIRAGE, which anchors feed-forward visual geometry to LiDAR--IMU metric motion and regulates online pose corrections according to correspondence support and cross-modal agreement, jointly enabling accurate metric pose estimation and high-fidelity Gaussian reconstruction. Experiments on Oxford Spires, M2DGR and INS show that MIRAGE achieves accurate metric trajectory estimation, outperforming COLMAP, a widely used but computationally expensive reference for vision-only benchmarks. Motivated by the scarcity of long-horizon monocular benchmarks for large-scale indoor reconstruction, we further build a multimodal data-collection platform and use MIRAGE to establish accurate 6-DoF reference trajectories for a new vision-only benchmark. We benchmark recent monocular large-scale reconstruction methods on this dataset, providing a systematic evaluation of their performance in long-horizon indoor scenes.

Xinzhe Wang, Changjian Jiang, Kaiwen Song +8

Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · The University of Hong Kong +4

Feed-Forward 3D ReconstructionCamera Pose EstimationIndoor Scene Reconstruction3D Gaussian Splatting
AskPDF