SLAM Dunk: A Recipe for Long-Horizon Egocentric SLAM in the Wild

· arXiv ·

5 min read Original article ↗

Imagine wearing a headset with cameras and inertial sensors for a long, unplanned walk. The system must keep an accurate trajectory the whole way, because accumulated drift makes the resulting map and your location less useful. A pipeline that assumes its sensors behave identically from the first minute to the last can let small mismatches pile up. SLAM Dunk keeps a conventional offline visual–inertial pipeline but lets the stereo-camera calibration change over time and corrects the gyroscope's response while it refines the path. In the authors' analysis of four training walks, error per distance walked stays roughly level as recordings get longer when the camera model is constant, while it falls toward the reference's stated accuracy when the camera model is allowed to vary. This is an offline result on selected LaMAria walks, not a guarantee for every trajectory.

Error per kilometre versus recording length for constant and time-varying calibration
Four LaMAria training walks were re-solved as 84 recordings of 2–20 minutes plus the full walks. With constant calibration (orange), error per kilometre stays comparatively flat as length grows; with time-varying calibration (blue), it declines. The fitted floors are 0.77 m/km (95% CI 0.48–1.04) for constant calibration and 0.09 m/km (95% CI 0.00–0.20) for time-varying calibration, the latter indistinguishable from zero at the reference's roughly 20 cm accuracy.

The floor result is reported in the paper's Figure 1.

The core machinery is bundle adjustment: jointly refining a trajectory and 3D landmarks so they agree with camera observations and inertial measurements. The baseline is a conventional offline pipeline with a RootSIFT front end. SLAM Dunk changes two things inside it.

First, camera drift. Time is divided into 5-second chunks. For each chunk the system estimates the rotation and translation of camera 1 relative to camera 0, plus both cameras' focal lengths. Successive chunk estimates are tied together by a random-walk prior, so calibration can wander but not jump arbitrarily. Principal points and distortion stay at factory values, and the transform from the stereo reference frame to the IMU is held fixed to limit ambiguity.

Second, gyroscope error. The system fits one global full 3×3 correction matrix capturing scale errors and cross-axis misalignment. Small response errors like these matter most during fast head rotation.

The paper offers evidence consistent with physical sensor effects. Fitted focal lengths rise with IMU temperature in 22 of 23 recordings for camera 0 and 23 of 23 for camera 1. The symmetric part of the fitted gyro correction has median cosine similarity 0.85 within devices and −0.21 between devices, across 23 recordings from four devices. These are correlations and repeatability checks, not proof that temperature alone causes the fitted changes.

The skeptical question is whether gains come from these calibration changes or from other pipeline differences. A factorial experiment keeps the surrounding system fixed and crosses constant versus time-varying camera calibration with the gyro correction either held at zero or fitted.

Mean CP score by gyro setting for constant and time-varying camera calibration
Figure 4(a), the CP-score panel: mean CP (higher is better) for gyro setting G=0 versus G fitted. Constant camera calibration (orange) gives 69.1 and 71.5; time-varying calibration (blue) gives 77.1 and 82.4. Fitting G adds +2.33 CP with constant calibration and +5.29 with time-varying calibration. Training results on 13 recordings with control points.

Fitting the gyro correction adds 2.3 CP with a constant camera calibration and 5.3 CP with time-varying calibration; the camera path adds 8.0 CP without the gyro correction and 11.0 with it. The gains are complementary in this pipeline, which does not show that the modeled phenomena are the sole cause of improvement. Here CP is the control-point score defined below.

On the training set, the two additions together raise mean CP from 69.1 to 82.4 (+13.3, 95% bootstrap CI 6.8–20.3) on 13 recordings with surveyed control points. On 21 recordings with dense references, mean absolute trajectory error (ATE, lower is better) falls from 1.02 m to 0.43 m, a 58% reduction. The CP gain occurs on 11 of 13 recordings and the ATE reduction on 16 of 21. The interaction between the additions is +3.0 CP (CI 0.6–5.7), positive on 11 of 13. One recording, 1_20, drops by 4.8 CP, and neither configuration solves 5_12. These comparisons isolate the camera and gyro toggles within one pipeline; they do not isolate every possible cause.

The held-out test is the LaMAria leaderboard, with five categories: Short, Medium, Long, Low light, and Moving. CP works as follows: triangulate surveyed control points, align them to surveyed positions with a robust similarity transform, then score each point's horizontal error on a piecewise-linear scale (100 through 5 cm, 60 at 1 m, 20 at 5 m, 0 at 10 m). Untriangulated points score zero, and the reported score averages points. In Medium, SLAM Dunk scored 78.9 versus Aria's 78.5. It exceeded Aria in Medium, Long, and Low light, but not Short or Moving. Aria stays slightly ahead on averages: 75.86 versus 75.46 for the mean of the five category scores weighted equally, and 77.13 versus 76.68 for the mean across recordings, where category counts differ. The snapshot is dated 3 November 2026 with 28 complete entries. This is near-parity on this benchmark, not an overall win.

Much of the idea is not new in kind: the LaMAria paper (arXiv) already describes long egocentric recordings and temperature-related focal-length changes, and the Project Aria report (arXiv) describes a proprietary service estimating time-varying camera and IMU calibration. SLAM Dunk contributes an open formulation within conventional bundle adjustment that measures the effect of each addition.

SLAM Dunk uses non-causal, offline optimization, so it is unsuitable for real-time use as presented. Its bundle adjustment took a median of 2 minutes and at most 5.5 minutes per training recording on six CPU threads, excluding seed VIO and feature extraction.

Within this pipeline, letting calibration vary is not just a better description of the sensors: the controlled comparison shows it improves trajectories, and the held-out result shows it brings an open method to roughly Aria's level on this benchmark.