Skip to content

LaMAria City-Scale Visual-Inertial SLAM Benchmark

Related docs: SLAM Benchmarking Metrics and Datasets, OpenVINS, VINS-Mono and VINS-Fusion, ORB-SLAM2 and ORB-SLAM3, MASt3R-SLAM, and SLAM3R / VGGT Foundation SLAM.

Last updated: 2026-05-23

Executive Summary

LaMAria is an ICCV 2025 egocentric visual-inertial SLAM dataset and benchmark from ETH Zurich, Meta Reality Labs Research, Google, and Microsoft Spatial AI Lab. It was built to test a failure mode that classic EuRoC/TUM-style VIO benchmarks do not cover well: long, city-scale wearable motion with dynamic visual content, low-light sections, vehicle or moving-platform segments, and time-varying calibration.

The key value is not that LaMAria looks like a road AV sensor suite. It does not. Its value is that it gives SLAM researchers a much harder visual-inertial metric benchmark with control-point based ground truth at city scale. For AV, warehouse, campus, and airside readers, it is a stress test for visual-inertial fallback and map-inspection pipelines, not a replacement for LiDAR-inertial or scan-to-map localization validation.

What It Contains

The official dataset page describes:

  • 23 training sequences and 63 test sequences.
  • Raw Aria recordings, Aria calibration, pinhole-converted ASL format data, ROS 1 bags, and pinhole calibration files.
  • Sparse control-point ground truth and pseudo-dense ground truth for training sequences.
  • Closed test-set ground truth with benchmark submission through the website.
  • Sequence categories spanning controlled experiments, additional short/medium/long routes, low-light routes, and moving-platform routes.
  • Code under the public cvg/lamaria repository for downloading and working with the dataset.

The paper frames the benchmark around glasses-like egocentric sensors rather than vehicle-mounted cameras. That matters: wide head/body motion, short-range visual clutter, occlusions, and moving people create different failure patterns from a rigid AV roof sensor bar.

Evaluation Model

LaMAria exposes two complementary evaluation styles:

Evaluation signalWhat it measuresWhy it matters
Sparse control-point alignmentDrift against centimeter-accurate surveyed control pointsAllows long trajectory evaluation without dense ground truth everywhere.
Dense pseudo-ground-truth pose recallFraction of keyframes within a position-error threshold after sparse alignmentCatches failures over long city routes, including low-light and moving-platform sections.
Benchmark tracksDifferent difficulty levels and held-out test dataReduces overfitting and makes immature systems easier to diagnose.
Calibration variantsRaw device model plus pinhole-converted dataSeparates algorithm limitations from camera-model support limitations.

For this corpus, report LaMAria results separately from EuRoC, TUM VI, KITTI, and Hilti-style construction benchmarks. The domain, camera geometry, and motion profile are different enough that averaging scores would hide the actual failure mode.

Domain Fit

DomainFitNote
Visual-inertial SLAM researchStrongDirectly tests city-scale egocentric VIO/SLAM under dynamic content and long routes.
Indoor service robots / warehousesConditionalUseful for low-light, dynamic, and calibration stress; not a floor-plan or wheel-odometry benchmark.
Outdoor campus robotsConditionalLong routes and lighting variation transfer; wearable motion and sensor placement differ.
Road AVWeak to conditionalUseful for camera/IMU fallback studies and visual map QA, but not a vehicle sensor-suite benchmark.
AirsideWeak to conditionalGood for visual-inertial failure analysis; does not test aircraft, wet apron, repeated stands, LiDAR maps, RTK, or radar.

Failure Modes It Exposes

  • Visual-inertial tracking loss under low light and dynamic visual clutter.
  • Scale or drift accumulation over long routes when metric constraints are weak.
  • Calibration sensitivity for unusual camera models and time-varying device behavior.
  • Overconfident pose estimates that look smooth locally but diverge at control points.
  • Benchmark overfitting when only a small indoor VIO dataset is used.

Implementation Notes

  • Treat the closed test ground truth as part of the benchmark contract; do not tune directly on test feedback.
  • Preserve the distinction between raw Aria data and pinhole-converted data when comparing methods.
  • Report whether a method uses IMU, map priors, learned depth, retrieval, or external scale cues.
  • Compare against OpenVINS/VINS/ORB-style classical baselines and foundation-model visual SLAM baselines separately.
  • Use LaMAria as a visual-inertial stress benchmark beside, not instead of, LiDAR/radar/RTK benchmarks for AV localization.

Limitations

  • It is an egocentric wearable dataset, not a vehicle-mounted AV dataset.
  • It does not validate 3D LiDAR map construction, scan-to-map localization, radar localization, or vehicle dynamics.
  • Test-set ground truth is intentionally closed for benchmark integrity.
  • Large downloads and format variants make reproducibility dependent on exact preprocessing.
  • It should not be used to claim airside or road AV readiness without target-domain data.

Sources

Public research notes collected from public sources.