Active Illumination Control in Low-Light Environments using NightHawk

Karthik Dantu; Yash Turkar; Youngjin Kim

REVIEW 3 major objections 5 minor 26 references

Active Illumination Control in Low-Light Environments using NightHawk

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read NightHawk jointly optimizes an onboard LED's brightness and the camera's exposure to keep visual features detectable and matchable in dark, texture-poor environments, with field tests reporting 47–197% better feature tracking than…

desk verdict NightHawk's field results contradict its own utility metric, and the paper lacks the statistical support for its headline numbers, but the idea is novel and the deployment is real. read the letter →

arxiv 2506.06394 v1 pith:XYA7RLHN submitted 2025-06-05 cs.RO cs.CV

Yash Turkar , Youngjin Kim , Karthik Dantu This is my paper

classification cs.ROcs.CV

keywords activeilluminationexposurecontrolBayesianoptimizationlow-lightrobotvisionfeaturedetectionandmatchingimageutilitymetricR2D2subterraneaninspection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

NightHawk claims that robot vision in dim, feature-poor environments improves when camera exposure time and an onboard LED's intensity are optimized together rather than handled separately by auto-exposure or a fixed light. The paper introduces an image utility score built from a learning-based feature detector's per-pixel repeatability and reliability, and uses that score as the objective for an online Bayesian optimizer that selects light intensity and exposure time as the robot moves. In field runs inside a 66-meter culvert beneath the Erie Canal, NightHawk produced average feature track lengths of 6.94 frames, compared with 2.33 for auto-exposure alone and 4.70 for a fixed 100% LED, corresponding to a 47–197% improvement in feature tracking. If the claim is right, inspection robots in tunnels, pipes, and other low-light structures can keep visual odometry and mapping reliable while drawing less power than constant illumination.

What carries the argument

The load-bearing object is the image utility metric $M_{\text{feat}}$, a scalar formed by multiplying the mean repeatability of R2D2, a learned feature detector, by the square of its mean reliability over all pixels. This converts the detector's per-pixel confidence maps into a single number that the optimizer can maximize. The optimizer is a Gaussian process with a Matérn kernel and Expected Improvement acquisition, and it runs inside an event-triggered loop: NightHawk keeps the current best configuration $(\Delta t^*, P^*)$ and restarts optimization only when $M^*_{\text{feat}} - M_{\text{feat}} > \epsilon$ for a user-set tolerance. The same network pass that estimates $M_{\text{feat}}$ can also extract R2D2 features, which is what lets the metric double as both quality assessment and feature extraction at 15 Hz on the onboard computer.

What would settle it

Run a controlled sweep of light intensity and exposure time in a low-light reflective tunnel and compare $M_{\text{feat}}$ with the number of features actually matched between consecutive frames; if the configuration that maximizes $M_{\text{feat}}$ does not also maximize the match count or track length, NightHawk's utility metric and its optimization target are misaligned with the claimed improvement.

Watch

Extended reading notes

Core claim

NightHawk's central claim is that active illumination and exposure control should be solved as one joint optimization, using a learned feature detector's confidence maps as the objective. The system maximizes $M_{\text{feat}} = \left(\frac{1}{N}\sum_{i=1}^{N} R_i\right)\left(\frac{1}{N}\sum_{i=1}^{N} Q_i\right)^2$, where $R_i$ is R2D2's per-pixel repeatability and $Q_i$ its per-pixel reliability. A Gaussian-process surrogate with a Matérn kernel and Expected Improvement selects the next light intensity $P$ and exposure time $\Delta t$, and an event trigger re-runs the search whenever the live metric drops more than a user-set tolerance below the current best. In the culvert experiments, NightHawk kept $M_{\text{feat}}$ from collapsing at the entrance and exit transitions, yielding average feature track lengths—how many consecutive frames a matched feature survives—of 6.94 frames versus 2.33 and 4.70 for the two baselines, with exposure time reduced from 14.98 ms to 7.77 ms.

Load-bearing premise

The paper assumes that R2D2's learned per-pixel repeatability and reliability values, averaged into $M_{\text{feat}}$, remain meaningful indicators of real feature detection and matching quality in the dim, reflective, texture-sparse culvert; if those confidence maps do not transfer to this domain, the Bayesian optimizer is maximizing a surrogate disconnected from the actual goal.

Editorial extensions

If this is right

Longer and steadier feature tracks should translate into more robust visual odometry, mapping, and inspection in pipes, tunnels, and other low-light structures.
Constraining exposure time by the desired frame rate keeps NightHawk's solutions within real-time limits while avoiding motion blur.
Choosing 54% LED intensity in the culvert instead of 100% reduces power draw, so an inspection robot can extend battery life without sacrificing feature persistence.
Because $M_{\text{feat}}$ correlates with matching performance across several detectors, a single computation can serve both image-quality assessment and feature extraction when R2D2 is used.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the authors do not develop is to run the same $M_{\text{feat}}$-guided optimization over exposure and gain only, without an active light, which could improve standard cameras in low light at lower hardware cost.
The event-triggered stop-and-go routine could be replaced by a predictive model that anticipates lighting transitions before the robot enters a dark zone, making the system feedforward rather than reactive.
The metric's dependence on a pretrained detector means its transfer to other domains is an open question: swapping R2D2 for a detector trained on similar low-light data would be a direct test of how general NightHawk's optimization target really is.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: NightHawk is a genuine systems contribution—jointly optimizing an external LED and exposure time under Bayesian optimization, with a utility metric built from R2D2's repeatability/reliability maps. That combination is new, and the field deployment on a Spot robot inside an Erie Canal culvert is the real thing. The paper is worth refereeing, but the empirical case is weaker than the abstract suggests.

What's actually good: the framing is practical, the event-triggered recursive BO pipeline is sensible, and the comparison against two baselines (AE with no light, AE with fixed full light) is the right kind of experiment. The Mfeat metric is a reasonable idea, and Figure 3 shows it correlates with matching across several detectors, with lower variance than gradient-based metrics. The authors also honestly note the compute advantage: one R2D2 pass can serve both quality assessment and feature extraction.

The soft spots are not minor. First, the deployment numbers come from a single run: average track lengths of 2.33, 4.70, and 6.94 appear with no variance, no trial count, and no raw logs. The abstract's '47-197%' is not backed by any uncertainty quantification. Second—and this is the load-bearing problem—the paper's own field data invert the utility metric. Section 4.1 says NightHawk runs at 54% light, gets 'slightly lower' Mfeat than the fixed-light baseline, yet achieves longer tracks. That means Mfeat, the thing the optimizer maximizes, is not tracking the outcome it is supposed to predict in the exact regime where the gain is claimed. Figure 3's correlation study doesn't resolve this; it isn't computed on these field frames or compared across the three control policies. Third, the R2D2-based metric is self-referential if the tracking evaluation also uses R2D2; Figure 3 includes other detectors, which helps, but the field evaluation needs to be explicit.

These issues are fixable. The authors should release the deployment logs, report multiple runs with error bars, and either show that Mfeat and matching co-vary within the deployment or accept that the optimizer is essentially improving something else. With that, the claim becomes credible.

Bottom line: serious referee, yes—the novelty and real deployment justify the time. But the revision needs to address the metric inversion and the statistics. I'd bring it to reading group, and might cite it after the evidence firms up.

Referee Report

3 major / 5 minor

Summary. NightHawk is an active-illumination and exposure-control framework for low-light robot vision. The paper proposes an image-utility metric Mfeat (Eq. 1), formed by aggregating R2D2's per-pixel repeatability and reliability tensors, and uses it as the cost function for an event-triggered Bayesian optimizer that jointly selects external LED intensity P and exposure time Δt. The system is deployed on a Boston Dynamics Spot robot inspecting a culvert beneath the Erie Canal. The field evaluation compares three settings—auto-exposure with no light, auto-exposure with fixed 100% light, and NightHawk—and reports average feature track lengths of 2.33, 4.70, and 6.94 frames, respectively, along with qualitative images and an online Mfeat trace. The paper also reports a Spearman-correlation study comparing Mfeat with existing utility metrics across several feature detectors.

Significance. If the field results hold, NightHawk is a worthwhile practical contribution for inspection robots in subterranean environments: it combines a learned feature-detector-based utility metric with online Bayesian optimization of both illumination and exposure, and it was validated in a real 66-meter culvert rather than only in simulation. The idea of reusing R2D2's repeatability/reliability outputs as a perceptual quality signal is interesting and could be reused beyond this specific system. The genuine field deployment and the attention to event-triggered, computationally feasible online operation are strengths. However, the paper's central quantitative claim currently rests on a single set of mean track lengths with no variance, and the deployment results appear to invert the paper's own metric-outcome relationship; these issues need to be resolved before the contribution can be considered established.

major comments (3)

[§4.1, Figures 4 and 5] The deployment results invert the paper's own premise that Eq. (1) is a utility metric for feature matching. The text states that NightHawk's chosen setting (54% light) yields 'slightly lower scores as in Figure 5' than the AE & P = 1.0 baseline, while Figure 4 reports a longer average track (6.94 vs 4.70). Under the paper's own definition, a lower Mfeat should correspond to worse matching; without showing that Mfeat and feature-track length co-vary within the same deployment logs, the reported track improvement cannot be attributed to maximizing Eq. (1). The authors should provide a per-segment or per-image comparison of Mfeat versus matching performance from the field runs, ideally across the three control policies on the same frame intervals.
[§4.1, Figure 4] The headline quantitative claim rests on single-run mean track lengths (2.33, 4.70, and 6.94 frames) with no variance, no number of trials, and no statistical test. The abstract's '47-197%' is simply the ratio of these three means. The text says that 'several tele-operated missions' were conducted; please report per-run values, frame-level distributions or confidence intervals, and a bootstrap or permutation test to establish that the differences are not within-run noise.
[§2.1, §3, §4.2] The metric-to-outcome link is not established for the deployment domain. Figure 3's Spearman correlations are not described in terms of the image set used: the text does not state whether they are computed on culvert frames or under any of the three control policies, and the track-length evaluation in Figure 4 does not state which feature detector is used. If the tracking evaluation uses R2D2, then Mfeat—which is derived from R2D2's repeatability and reliability tensors—may be a self-fulfilling objective. If a different detector is used, the claim that maximizing R2D2-derived Mfeat improves general feature matching is unsupported. Please specify the detector used for the tracking evaluation and provide the correlation analysis on the field images themselves.

minor comments (5)

[§2.2, Eqs. (2)–(3)] The notation is confusing: x = [P, ΔT] is defined, but then 'inputs x* = {x0, . . . , xm}' uses x* both as a test input and as a set of training locations. Please rewrite with distinct symbols for training points, test points, and the control vector.
[Figure 4 caption] The caption contains unlabeled values ('= 7.77 ms', '= 14.98 ms', '= 2.33', etc.) and the legend is unclear; the reader cannot tell which axis or symbol corresponds to which configuration. Please add explicit labels and units.
[§4.1] The phrase 'good feature matching accuracy in our application' is asserted without a quantitative definition; the text should define the matching metric (number of matches, track length, or inlier ratio) consistently with Figure 4.
[Figure 3] The axis labels 'Feature ID' and 'Frame ID' appear to be swapped or missing units; please clarify what is being plotted and add error bars or variance information, since 'low variance' is claimed in the text.
[§4.3] The claim that 'multi-threading and early stopping' reduced optimization latency from 70 to 20 seconds is reported without an experimental protocol; please state the hardware, number of runs, and how the times were measured, or mark the number as illustrative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the NightHawk derivation is self-contained, and the metric/outcome gap is a validity concern rather than a circular one.

full rationale

The paper's chain is to define an image-utility metric Mfeat from R2D2 outputs (Eq. 1, Sec. 2.1), maximize that metric with Bayesian optimization (Sec. 2.2), and then report field feature-tracking gains (Fig. 4). No step equates Mfeat with the reported outcome: average track length is a separate empirical measurement, and Mfeat is not fitted to those tracks. The only self-referential element is that Mfeat aggregates R2D2's repeatability/reliability, and the paper notes a single R2D2 pass can provide both Mfeat and features (Sec. 3: 'A single computation for both feature extraction and quality assessment (e.g., when using R2D2 features) can help reduce compute overheads'). If the field tracker were known to be R2D2-based, some overlap between the optimized metric and R2D2 track persistence would be expected, but the paper validates Mfeat against AKAZE, Shi-Tomasi, ORB, and SuperPoint as well, giving independent content, and the field result is not shown to be R2D2-only. The self-citations (Refs. 19-20) concern frame-rate assumptions for SLAM services and are not load-bearing. The more serious issue is a validity gap, not circularity: Sec. 4.1 admits 'This results in slightly lower scores as in Figure 5, but good feature matching accuracy in our application,' so in the deployment Mfeat is lower than a baseline while track length is higher. That inversion means the improvement is not demonstrated to be caused by maximizing Eq. (1); however, because track length is not defined in terms of Mfeat, the derivation is not circular. Score 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central optimization objective Mfeat is built from a pretrained learned detector (R2D2) and two design choices (reliability square and user-set threshold), while the GP and acquisition function rely on standard smoothness and noise assumptions. These constitute the bulk of the unverified input the claim rests on.

free parameters (5)

Trigger threshold epsilon = not specified
User-provided tolerance that decides when re-optimization is triggered (Section 2.2). Its value affects how often NightHawk reconfigures.
Reliability exponent in Mfeat = 2
The square in Eq. (1) is chosen by hand to amplify reliability; no derivation or ablation is provided (Section 2.1).
GP kernel hyperparameters = not reported
Lengthscales and signal/noise variance for the Matérn kernel are not stated (Section 2.2), so the surrogate behavior is not reproducible.
Consecutive-frame trigger count = not reported
The design suppresses re-optimization unless the threshold is violated for a set number of consecutive frames; the count is not given (Section 4.1).
Search-space bounds for P and exposure time = not reported
The optimizer searches a user-defined range for light intensity and exposure time; bounds are not listed (Section 2.2).

assumptions (5)

domain assumption R2D2's pretrained repeatability and reliability tensors transferred from their training domain are meaningful proxies for feature quality in the dim, reflective culvert environment.
In Section 2.1, Mfeat is defined directly from R2D2 outputs; if the network's maps are not calibrated for this low-light domain, the optimizer maximizes a disconnected objective.
domain assumption The relationship between (P, exposure time) and Mfeat is smooth and well-approximated by a Gaussian process with Matérn kernel.
Used in Section 2.2 to justify the GP surrogate; no validation of the smoothness assumption is presented.
domain assumption Scene lighting is quasi-static during each 20-second optimization window.
Section 4.3 notes the optimizer can take 20 seconds and may pause the robot; if lighting changes significantly within that window, the optimal configuration found may be stale.
standard math Zero-mean Gaussian observation noise.
Standard assumption in the GP formulation of Section 2.2, with noise variance sigma_n^2.
domain assumption Feature track length is a valid measure of visual estimation reliability.
Used when comparing NightHawk to baselines in Figure 4; longer tracks are assumed to imply more reliable visual estimation, but no downstream SLAM or odometry error is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Active Illumination Control in Low-Light Environments using NightHawk." pith.science (2026). https://pith.science/paper/XYA7RLHN

@misc{pith2026250606394,
  author       = {Pith},
  title        = {Pith review of: Active Illumination Control in Low-Light Environments using NightHawk},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYA7RLHN}},
  note         = {Machine review of arXiv:2506.06394}
}

read the original abstract

Subterranean environments such as culverts present significant challenges to robot vision due to dim lighting and lack of distinctive features. Although onboard illumination can help, it introduces issues such as specular reflections, overexposure, and increased power consumption. We propose NightHawk, a framework that combines active illumination with exposure control to optimize image quality in these settings. NightHawk formulates an online Bayesian optimization problem to determine the best light intensity and exposure-time for a given scene. We propose a novel feature detector-based metric to quantify image utility and use it as the cost function for the optimizer. We built NightHawk as an event-triggered recursive optimization pipeline and deployed it on a legged robot navigating a culvert beneath the Erie Canal. Results from field experiments demonstrate improvements in feature detection and matching by 47-197% enabling more reliable visual estimation in challenging lighting conditions.

Figures

Figures reproduced from arXiv: 2506.06394 by the authors.

**Figure 1.** Active illumination and exposure control on Boston Dynamics Spot, inspecting culverts under the Erie Canal in Medina, NY. (Left) Spot approaches entrance of culvert 110, a 66m long, subterranean environment with extremely low ambient illumination. (Right) Spot entering and exiting the culvert equipped with NightHawk 1 This project was partially funded by NSF #1846320 and a gift from MOOG Inc. arXiv:2506.06394v1 [c… view at source ↗

**Figure 2.** Overview of the NightHawk pipeline NightHawk’s overall architecture is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Correlation of utility metrics with feature detection/matching. Feature ID Frame ID NightHawk = 14.98 ms = 14.98 ms = 7.77 ms = 2.33 = 4.70 = 6.94 AE & P = 0.0 AE & P = 1.0 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Feature tracking performance of the 3 settings where NightHawk shows improved feature tracking and lower exposure-times offers a significant practical advantage when using learning-based features in visual estimation. A single computation for both feature extraction …

**Figure 5.** Figure 5: Change in Mfeat as the robot enters the culvert is shown in the 3 configurations, AE with no external light results in underexposed images while adding a fixed (100%) light introduces artifacts such as a green-ish hue. Finally, with NightHawk, overall image utility is …

**Figure 6.** Figure 6: Deployment of our active illumination system in subterranean culvert environments. The top row shows the robot equipped with NightHawk approaching and entering a culvert beneath the Erie Canal. The bottom row illustrates the visual challenges encountered inside the c…

**Figure 7.** Figure 7: NightHawk system in action. (Left) Hardware prototype featuring a high-power 50W LED, stereo FLIR BlackFly S cameras, and an onboard ROS-based LED controller. (Right) Real-time response of the system during a live experiment, where Mfeat drops below a threshold, trigg…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

[1]

Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces

Pablo Alcantarilla, Jesus Nuevo, and Adrien Bartoli. Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces. InProcedings of the British Ma- chine Vision Conference 2013, pages 13.1–13.11, Bristol, 2013. British Machine Vision Association

work page 2013
[2]

SuperPoint: Self- Supervised Interest Point Detection and Description, April 2018

Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self- Supervised Interest Point Detection and Description, April 2018. arXiv:1712.07629 [cs]

arXiv 2018
[3]

Kamak Ebadi, Lukas Bernreiter, Harel Biggie, Gavin Catt, Yun Chang, Arghya Chatterjee, Christopher E. Denniston, Simon-Pierre Deschênes, Kyle Harlow, Shehryar Khattak, Lucas Nogueira, Matteo Palieri, Pavel Petráček, Matěj Petr- lík, Andrzej Reinke, Vít Krátký, Shibo Zhao, Ali-akbar Agha-mohammadi, Kostas Alexis, Christoffer Heckman, Kasra Khosoussi, Navin...

work page 2024
[4]

SiLK -- Simple Learned Keypoints

Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK – Simple Learned Keypoints, April 2023. arXiv:2304.06194 [cs]

work page Pith review arXiv 2023
[5]

Learning-Based Image Enhancement for Visual Odometry in Challenging HDR Environments

Ruben Gomez-Ojeda, Zichao Zhang, Javier Gonzalez-Jimenez, and Davide Scara- muzza. Learning-Based Image Enhancement for Visual Odometry in Challenging HDR Environments. In2018 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 805–811, Brisbane, QLD, May 2018. IEEE

work page 2018
[6]

Camera Attributes Control for Visual Odometry With Motion Blur Awareness

Bin Han, Yicheng Lin, Yan Dong, Hao Wang, Tao Zhang, and Chengyuan Liang. Camera Attributes Control for Visual Odometry With Motion Blur Awareness. IEEE/ASME Transactions on Mechatronics, 28(4):2225–2235, August 2023

work page 2023
[7]

Good features to track

Jianbo Shi and Tomasi. Good features to track. InProceedings of IEEE Conference on Computer Vision and Pattern Recognition CVPR-94, pages 593–600, Seattle, WA, USA, 1994. IEEE Comput. Soc. Press

work page 1994
[8]

Exposure Control Using Bayesian Optimization Based on Entropy Weighted Image Gradient

Joowan Kim, Younggun Cho, and Ayoung Kim. Exposure Control Using Bayesian Optimization Based on Entropy Weighted Image Gradient. In2018 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 857–864, Brisbane, QLD, May 2018. IEEE

work page 2018

Show all 26 references

[9]

Proactive Camera Attribute Con- trol Using Bayesian Optimization for Illumination-Resilient Visual Navigation

Joowan Kim, Younggun Cho, and Ayoung Kim. Proactive Camera Attribute Con- trol Using Bayesian Optimization for Illumination-Resilient Visual Navigation. IEEE Transactions on Robotics, 36(4):1256–1271, August 2020

2020
[10]

Learning to Control Camera Exposure via Reinforcement Learning

Kyunghyun Lee, Ukcheol Shin, and Byeong-Uk Lee. Learning to Control Camera Exposure via Reinforcement Learning. In2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 2975–2983, Seattle, WA, USA, June 2024. IEEE

2024
[11]

Xiadong Liu, Zhi Gao, Huimin Cheng, Pengfei Wang, and Ben M. Chen. Learning- based Low Light Image Enhancement for Visual Odometry. In2020 IEEE 16th International Conference on Control & Automation (ICCA), pages 1143–1148, Sin- gapore, October 2020. IEEE

2020
[12]

Robot Operating System 2: Design, architecture, and uses in the wild

Steven Macenski, Tully Foote, Brian Gerkey, Chris Lalancette, and William Woodall. Robot Operating System 2: Design, architecture, and uses in the wild. Science Robotics, 7(66):eabm6074, May 2022. Publisher: American Association for the Advancement of Science

2022
[13]

Ishaan Mehta, Mingliang Tang, and Timothy D. Barfoot. Gradient-Based Auto- Exposure Control Applied to a Self-Driving Car. In2020 17th Conference on Computer and Robot Vision (CR V), pages 166–173, Ottawa, ON, Canada, May
[14]

PyTorch: An Imperative Style, High-Performance Deep Learning Li- brary

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Des- maison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, L...
[15]

Scikit-learn: Machine Learning in Python.MACHINE LEARNING IN PYTHON

Fabian Pedregosa, Gael Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, and David Cournapeau. Scikit-learn: Machine Learning in Python.MACHINE LE...
[16]

Carl Edward Rasmussen and Christopher K. I. Williams.Gaussian processes for machine learning. Adaptive computation and machine learning. MIT Press, Cam- bridge, Mass, 2006. OCLC: ocm61285753

2006
[17]

R2D2: Repeatable and Reliable Detec- tor and Descriptor, June 2019

Jerome Revaud, PhilippeWeinzaepfel,César DeSouza,NoePion,Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2D2: Repeatable and Reliable Detec- tor and Descriptor, June 2019. arXiv:1906.06195 [cs]. 12 Turkar et al

2019 arXiv
[18]

ORB: An effi- cient alternative to SIFT or SURF

Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An effi- cient alternative to SIFT or SURF. In2011 International Conference on Computer Vision, pages 2564–2571, Barcelona, Spain, November 2011. IEEE

2011
[19]

A Comprehensive Study of Systems Challenges in Visual Simultaneous Localiza- tion and Mapping Systems.ACM Transactions on Embedded Computing Systems, 24(1):1–31, January 2025

Sofiya Semenova, Steven Ko, Yu David Liu, Lukasz Ziarek, and Karthik Dantu. A Comprehensive Study of Systems Challenges in Visual Simultaneous Localiza- tion and Mapping Systems.ACM Transactions on Embedded Computing Systems, 24(1):1–31, January 2025

2025
[20]

Ko, Yu David Liu, Lukasz Ziarek, and Karthik Dantu

Sofiya Semenova, Steven Y. Ko, Yu David Liu, Lukasz Ziarek, and Karthik Dantu. A quantitative analysis of system bottlenecks in visual SLAM. InProceedings of the 23rd Annual International Workshop on Mobile Computing Systems and Applications, pages 74–80, Tempe Arizona, March ...

2022
[21]

Auto-adjusting camera exposure for outdoor robotics using gradient information

Inwook Shim, Joon-Young Lee, and In So Kweon. Auto-adjusting camera exposure for outdoor robotics using gradient information. In2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1011–1017, Chicago, IL, USA, September 2014. IEEE

2014
[22]

Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. Practical Bayesian Optimiza- tion of Machine Learning Algorithms, August 2012. arXiv:1206.2944 [stat]

2012 arXiv
[23]

Waslander, and Jonathan Kelly

Justin Tomasi, Brandon Wagstaff, Steven L. Waslander, and Jonathan Kelly. Learned Camera Gain and Exposure Control for Improved Visual Feature De- tection and Matching.IEEE Robotics and Automation Letters, 6(2):2028–2035, April 2021

2021
[24]

An Image Acquisition Scheme for Visual Odometry based on Image Bracketing and Online Attribute Control

Shuyang Zhang, Jinhao He, Bohuan Xue, Jin Wu, Pengyu Yin, Jianhao Jiao, and Ming Liu. An Image Acquisition Scheme for Visual Odometry based on Image Bracketing and Online Attribute Control. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 381–387, ...

2024
[25]

Efficient Camera Exposure Control for Visual Odometry via Deep Reinforcement Learning.IEEE Robotics and Automation Letters, 10(2):1609–1616, February 2025

Shuyang Zhang, Jinhao He, Yilong Zhu, Jin Wu, and Jie Yuan. Efficient Camera Exposure Control for Visual Odometry via Deep Reinforcement Learning.IEEE Robotics and Automation Letters, 10(2):1609–1616, February 2025

2025
[26]

Active exposure con- trol for robust visual odometry in HDR environments

Zichao Zhang, Christian Forster, and Davide Scaramuzza. Active exposure con- trol for robust visual odometry in HDR environments. In2017 IEEE International Conference on Robotics and Automation (ICRA), pages 3894–3901, Singapore, Sin- gapore, May 2017. IEEE

2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces

Pablo Alcantarilla, Jesus Nuevo, and Adrien Bartoli. Fast Explicit Diffusion for Accelerated Features in Nonlinear Scale Spaces. InProcedings of the British Ma- chine Vision Conference 2013, pages 13.1–13.11, Bristol, 2013. British Machine Vision Association

work page 2013

[2] [2]

SuperPoint: Self- Supervised Interest Point Detection and Description, April 2018

Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self- Supervised Interest Point Detection and Description, April 2018. arXiv:1712.07629 [cs]

arXiv 2018

[3] [3]

Kamak Ebadi, Lukas Bernreiter, Harel Biggie, Gavin Catt, Yun Chang, Arghya Chatterjee, Christopher E. Denniston, Simon-Pierre Deschênes, Kyle Harlow, Shehryar Khattak, Lucas Nogueira, Matteo Palieri, Pavel Petráček, Matěj Petr- lík, Andrzej Reinke, Vít Krátký, Shibo Zhao, Ali-akbar Agha-mohammadi, Kostas Alexis, Christoffer Heckman, Kasra Khosoussi, Navin...

work page 2024

[4] [4]

SiLK -- Simple Learned Keypoints

Pierre Gleize, Weiyao Wang, and Matt Feiszli. SiLK – Simple Learned Keypoints, April 2023. arXiv:2304.06194 [cs]

work page Pith review arXiv 2023

[5] [5]

Learning-Based Image Enhancement for Visual Odometry in Challenging HDR Environments

Ruben Gomez-Ojeda, Zichao Zhang, Javier Gonzalez-Jimenez, and Davide Scara- muzza. Learning-Based Image Enhancement for Visual Odometry in Challenging HDR Environments. In2018 IEEE International Conference on Robotics and Au- tomation (ICRA), pages 805–811, Brisbane, QLD, May 2018. IEEE

work page 2018

[6] [6]

Camera Attributes Control for Visual Odometry With Motion Blur Awareness

Bin Han, Yicheng Lin, Yan Dong, Hao Wang, Tao Zhang, and Chengyuan Liang. Camera Attributes Control for Visual Odometry With Motion Blur Awareness. IEEE/ASME Transactions on Mechatronics, 28(4):2225–2235, August 2023

work page 2023

[7] [7]

Good features to track

Jianbo Shi and Tomasi. Good features to track. InProceedings of IEEE Conference on Computer Vision and Pattern Recognition CVPR-94, pages 593–600, Seattle, WA, USA, 1994. IEEE Comput. Soc. Press

work page 1994

[8] [8]

Exposure Control Using Bayesian Optimization Based on Entropy Weighted Image Gradient

Joowan Kim, Younggun Cho, and Ayoung Kim. Exposure Control Using Bayesian Optimization Based on Entropy Weighted Image Gradient. In2018 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 857–864, Brisbane, QLD, May 2018. IEEE

work page 2018