PaperlayerView exact revision on arXiv

Source-linked AI summary

Bracketing Shows the Way: Adaptive Multi-Exposure Policy as Input for Robust Visual SLAM

Olivier Gamache, Anthony Lavertu, David-Alexandre Duclos, Jean-Michel Fortin, François Pomerleau, Philippe Giguère

arXiv:2610.33174v1cs.RO

TL;DR

Rapid illumination changes can saturate VSLAM images before automatic exposure adapts, removing visual constraints. The paper pairs adaptive exposure bracketing with a factor-graph estimator that directly uses bracketed images. In the tunnel-exit disturbance, the combination reduces state-estimation error by 78% relative to AE images.

  • Problem

    Reactive AE can adapt too slowly during rapid illumination changes, saturating images and removing the visual constraints VSLAM needs for tracking.

    Source: §I · p. 1

  • Method

    B-Adapt adapts exposure brackets to cover scene dynamic range, while BVSLAM directly incorporates the resulting multi-exposure images as factor-graph constraints without HDR merging.

    Source: §I · pp. 1–2, §VII · p. 7

  • Results

    78% reduction in state-estimation error relative to AE was achieved by BVSLAM with B-Adapt on the tunnel trajectory.

    Source: A. VSLAM Analysis on Nominal and… · p. 6

  • Takeaways & Limitations

    The B-Adapt and BVSLAM combination performs on par with or better than evaluated state-of-the-art approaches across nominal and disturbance conditions, with the lowest tunnel-exit error.

    Source: §VII · p. 7

    BVSLAM requires offline processing because of its overall runtime and global bundle adjustment.

    Source: B. Limitations · p. 7

Abstract

from arXiv · show

Rapid illumination transitions can disturb Visual SLAM (VSLAM) by saturating images before the Automatic Exposure (AE) policy adapts, removing the visual correspondences needed to maintain tracking. Our proposed methods mitigate this disturbance at two levels: B-Adapt acquires adaptive exposure brackets to cover the scene dynamic range while promoting photometric overlap, and BVSLAM directly uses the multi-exposure sequence as constraints in a factor graph state estimator without relying on High Dynamic Range (HDR) merging. We evaluate both methods against state-of-the-art approaches in nominal environments and during a repeated tunnel-exit transition at an emulated vehicle speed of 32 m s-1. In the tunnel, combining our B-Adapt with BVSLAM reduces the state estimation error by 78 % compared with AE images. These results suggest that bracketing efficiently improves visual pose estimation robustness to rapid illumination disturbances while retaining low error under nominal conditions. Our code is available at https://github.com/norlab-ulaval/bracketing-shows-the-way.

I. INTRODUCTION

Rapid illumination changes can saturate VSLAM images before reactive automatic exposure adapts, removing visual constraints and potentially causing tracking failure. The paper proposes adaptive bracketing and direct factor-graph use of bracketed images as complementary mitigations.

I. INTRODUCTION

  • Rapid illumination transitions can saturate images during AE’s adaptation delay, depriving VSLAM of visual constraints and potentially causing tracking failure.At tunnel exits, scene dynamic range can reach 120 dB, versus a theoretical 48 dB for 8-bit acquisition.Source: §I · p. 1
  • Existing bracketing methods commonly merge exposures into HDR images, while current VSLAM systems are not designed to directly exploit brackets acquired under camera motion.Motion during HDR reconstruction can introduce ghosting and misregistration artifacts.Source: §I · p. 1
  • The paper introduces B-Adapt to cover scene dynamic range with adaptive brackets and BVSLAM to process bracketed images directly in a factor graph without HDR merging.It evaluates the two strategies against state-of-the-art approaches in nominal conditions and disturbance-affected regions.Source: §I · pp. 1–2

II. RELATED WORK

The related-work review covers image acquisition and pose-estimation approaches to illumination disturbance. Single-exposure policies adapt exposure selection, but remain limited by sensor dynamic range, motivating multi-exposure bracketing.

II. RELATED WORK

  • The review frames illumination-robust VSLAM through two complementary directions: image acquisition policies and state-estimation methods.It first surveys acquisition-side methods and then systems that incorporate multiple exposures into pose estimation.Source: §II · p. 2

A. Single-Exposure Acquisition for Robotics

  • Reactive exposure policies optimize from observed scenes and introduce delays, while predictive policies anticipate upcoming exposure needs.Regardless of selection strategy, these policies acquire one image at a time and remain limited by sensor dynamic range.Source: A. Single-Exposure Acquisition for… · p. 2
  • Because single-exposure policies remain constrained by sensor dynamic range, the review motivates the paper’s multi-exposure bracketing policy.Source: A. Single-Exposure Acquisition for… · p. 2

B. Multi-Exposure Bracketing under Motion

Multi-exposure work addresses motion-related artifacts and adapts exposure spacing, but existing approaches face constraints in merging, hardware, or direct use of brackets for VSLAM. The paper combines adaptive bracketing with bracket-aware state estimation.

B. Multi-Exposure Bracketing under Motion

  • Under motion, HDR merging can produce ghosting and misregistration, and wider exposure coverage increases time and viewpoint changes between brackets.HDRFlow accommodates larger displacements using optical flow, but occlusions remain challenging for image merging.Source: B. Multi-Exposure Bracketing under… · p. 2
  • Prior bracketing policies adapt exposure spacing or separation to scene conditions, while specialized sensors avoid sequential viewpoint changes but are not commonly available on robotic platforms.The paper instead uses standard camera hardware to acquire bracket sequences as direct VSLAM input.Source: B. Multi-Exposure Bracketing under… · p. 2

C. VSLAM under Challenging Illumination

  • Existing VSLAM methods address appearance changes or use specialized multi-exposure systems, but current systems are not designed for bracketed sequences acquired under camera motion.The proposed combination uses B-Adapt for dynamic-range coverage and cross-exposure photometric overlap, and BVSLAM for estimation without HDR merging.Source: C. VSLAM under Challenging Illumination · pp. 2–3

III. THEORY

B-Adapt estimates bounding exposures from observed images and inserts intermediate levels to cover scene radiance. It sequences adjacent exposure levels to promote photometric overlap while maintaining separation between the bounds.

III. THEORY

A. B-Adapt: Bracketing Acquisition Policy

  • Low and High controllers use separate one-sided intensity-tail objectives, with α setting the tolerated pixel fraction and β providing a margin from sensor limits.The Low controller focuses on bright regions; the High controller focuses on dark regions.Source: A. B-Adapt: Bracketing Acquisition… · p. 3
  • For n ≥3, B-Adapt inserts n−2 intermediate exposures that evenly divide the EV interval between the bounding brackets.The intermediate levels are computed from the latest Low and High exposures, and a minimum-separation constraint prevents the bounds from converging too closely.Source: A. B-Adapt: Bracketing Acquisition… · pp. 3–4
  • The acquisition sequence moves only between adjacent exposure levels, avoiding consecutive Low and High frames when n ≥3 to maximize photometric overlap.Source: A. B-Adapt: Bracketing Acquisition… · p. 4
  • B-Adapt preemptively covers multiple exposure levels to preserve usable visual information during difficult lighting, unlike conventional AE’s emphasis on image appearance under gradual changes.Source: A. B-Adapt: Bracketing Acquisition… · p. 4

B. BVSLAM Implementation

BVSLAM directly derives visual constraints from bracketed images, linking landmarks across exposure levels in a factor graph without image fusion. Its implementation uses a zero-motion prior and is intended for offline analysis rather than real-time operation.

B. BVSLAM Implementation

  • BVSLAM matches images across exposure levels within a bounded exposure step and lookback window, avoiding associations with large exposure differences.The lookback window spans a complete bracketing cycle.Source: B. BVSLAM Implementation · p. 4
  • Stereo and temporal correspondences constrain camera poses and landmark positions in a factor graph, while a zero-motion prior prevents trajectory propagation through visually weak regions.The nonlinear least-squares problem is solved incrementally over a sliding window and then refined over the complete graph by global bundle adjustment.Source: B. BVSLAM Implementation · p. 4
  • The current BVSLAM implementation is offline and does not operate in real time.Its primary objective is to evaluate how bracketing affects visual constraints, not real-time performance.Source: B. BVSLAM Implementation · p. 4

IV. METHODOLOGY

The methodology evaluates the proposed acquisition and state-estimation methods in nominal conditions and during a rapid illumination increase, which the paper terms the disturbance condition.

IV. METHODOLOGY

  • The evaluation covers both the proposed image-acquisition and state-estimation methods.The section proceeds from the acquisition platform and implementation parameters to recorded environments and then evaluation comparisons, references, and metrics.Source: §IV · p. 4
  • The two operating conditions are nominal conditions and a rapid illumination increase termed the disturbance condition.Source: §IV · p. 4
  • The methodology section presents the platform and parameters, recorded environments, and the compared methods, reference trajectories, and evaluation metrics.Source: §IV · p. 4

A. Experimental Setup

The experiments use a sensor-equipped stereo acquisition head and test nominal trajectories alongside a rapid tunnel-exit illumination transition. The disturbance setup reduces camera frame rate to emulate larger inter-frame motion at vehicle speed.

A. Experimental Setup

  • The acquisition head combines synchronized global-shutter stereo cameras with lidar, an IMU, and a luxmeter for reference trajectories and illumination measurements.B-Adapt exposure-time and gain bounds were selected to balance motion-blur reduction against image brightness.Source: A. Experimental Setup · p. 4
  • Figure 3 shows the acquisition platform and example frames across nominal and disturbance trajectories, with B-Adapt and B-Fixed exposures stacked and cropped.Source: Figure 3 · p. 5

B. Experiment Environments

  • Nominal Indoor and Forest data span 650 m and were recorded at about 2.3 m s−1 with a 20 Hz camera.Source: B. Experiment Environments · p. 5
  • The 55 m tunnel trajectory focuses on the rapid low-to-high illumination transition at the exit; 4 Hz acquisition yields inter-frame displacement equivalent to about 32 m s−1 at 30 Hz.The gradual high-to-low transition at the tunnel entry was not treated as a disturbance.Source: B. Experiment Environments · p. 5

C. Evaluation Modalities

The evaluation crosses acquisition policies with state estimators and compares their trajectories against an illumination-independent lidar-inertial reference. It reports pooled disturbance RPE RMSE and examines saturation and feature correspondences to interpret behavior.

C. Evaluation Modalities

  • Acquisition comparisons include single-exposure AE, HDRFlow merging B-Adapt brackets, and B-Fixed brackets, while state estimators include BVSLAM, ORB-SLAM3, CuVSLAM, and AirSLAM.HDRFlow can introduce reconstruction distortions and artifacts under large inter-frame motion.Source: C. Evaluation Modalities · p. 5
  • Each disturbance acquisition policy was recorded five times over the same trajectory, while each nominal trajectory was recorded once per policy.Source: C. Evaluation Modalities · p. 5
  • The lidar-inertial reference is independent of illumination transitions affecting visual estimators, and estimates are evaluated against it.Source: C. Evaluation Modalities · p. 5
  • Trajectory accuracy uses RPE over 5 m windows, with disturbance windows pooled across repetitions and summarized by RMSE.RMSE emphasizes large errors associated with severe trajectory deviations or tracking loss.Source: C. Evaluation Modalities · p. 5

V. RESULTS

  • Figure 4 organizes RPE RMSE by estimator and acquisition policy for nominal conditions and the tunnel-exit disturbance; disturbance results pool five trajectories.The Results section also examines image saturation and feature correspondences to explain disturbance behavior.Source: §V · p. 5, Figure 4 · p. 6

A. VSLAM Analysis on Nominal and Disturbance Datasets

B-Adapt with BVSLAM achieves the lowest reported tunnel-trajectory error while retaining nominal performance. At the tunnel exit, B-Adapt preserves visual information and its trajectory deviates only slightly, unlike the AE-based estimate, which loses tracking.

A. VSLAM Analysis on Nominal and Disturbance Datasets

  • Under nominal conditions, all three estimators perform similarly with AE, while fixed bracketing raises CuVSLAM and AirSLAM errors because widely separated exposures hinder association.BVSLAM retains same-order error with fixed brackets, and adaptive spacing keeps all three estimators' nominal errors low.Source: A. VSLAM Analysis on Nominal and… · p. 6
  • BVSLAM with B-Adapt reaches 1.10 m error, a 78 % reduction relative to AE at 4.97 m, and beats B-Fixed at 3.73 m and HDRFlow at 2.91 m.The B-Adapt result remains comparable to the same combination's nominal performance.Source: A. VSLAM Analysis on Nominal and… · p. 6

B. Exposure Policies in Rapid Illumination Transition

  • At the tunnel exit, AE images become almost entirely overexposed and high-confidence matches fall nearly to zero for about one second, while B-Adapt's Low bracket retains observable content.B-Adapt has a brief adaptation delay, but informative correspondences remain available through the transition.Source: B. Exposure Policies in Rapid… · p. 6
  • The AE-based BVSLAM trajectory loses track near the tunnel exit, whereas the B-Adapt trajectory shows only a small deviation, as illustrated by the trajectory comparison.Figure 5 shows AE and B-Adapt trajectories alongside the reference trajectory.Source: B. Exposure Policies in Rapid… · p. 6, Figure 5 · p. 7
  • The combined method retains comparable nominal performance while substantially outperforming evaluated alternatives during the tunnel-exit disturbance at 32 m s−1.The comparison covers nominal and disturbance conditions.Source: B. Exposure Policies in Rapid… · p. 7

VI. DISCUSSION

B-Adapt with BVSLAM improves robustness to rapid illumination transitions by exploiting the complete exposure cycle. As emulated speed rises, AE errors increase sharply while B-Adapt remains comparatively stable.

VI. DISCUSSION

  • B-Adapt with BVSLAM improves robustness to rapid illumination transitions by exploiting the complete exposure cycle.The study next examines how the benefit changes with emulated vehicle speed.Source: §VI · p. 7

A. Effect of Emulated Speed on VSLAM Performance

  • Across evaluated frame-rate settings, B-Adapt performs at least as well as AE, and its advantage grows at higher emulated speeds as AE's adaptation delay becomes more consequential.The figure reports RPE RMSE pooled across five repeated trajectories per policy and frame-rate setting.Source: A. Effect of Emulated Speed on VSLAM… · p. 7

B. Limitations

B-Adapt retains usable visual information during the evaluated transition, but image-driven control can delay its response. BVSLAM currently requires offline processing, and future work targets faster adaptation and lower computational cost.

B. Limitations

  • B-Adapt's image-driven controllers cannot respond immediately, and an extreme-intensity region may be too small initially to shift the bounding bracket controller.The minimum bracket separation anticipates sharp transitions, while tuning α, β, and λ trades faster adaptation against policy stability.Source: B. Limitations · p. 7
  • BVSLAM requires offline processing because of its runtime and global bundle adjustment; real-time implementation remains a future direction.Future work also targets improved B-Adapt responsiveness while maintaining controller stability.Source: B. Limitations · p. 7, §VII · p. 8

VII. CONCLUSION

  • The proposed combination performs on par with or better than evaluated state-of-the-art approaches across nominal and disturbance conditions, with the lowest error during the tunnel exit.Its advantage becomes more pronounced at higher emulated speeds as AE's adaptation delay becomes more consequential.Source: §VII · pp. 7–8

Related papers

Loading 2610.33174v1…