Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping

Hashan K. Weerasooriya; Stanley Chan; Weijian Zhang

REVIEW 3 major objections 6 minor 29 references

Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping

T0 review · 3 major / 6 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 The paper's core claim is that dead-time-distorted single-photon LiDAR registration statistics can be predicted by a learned autoencoder mapping, making high-flux simulation dramatically faster without sacrificing statistical accuracy.

desk verdict A modest but useful neural surrogate for high-flux single-photon LiDAR simulation, with accuracy claims that are under-validated because the test RMSE is measured against noisy 20-realization Monte Carlo labels. read the letter →

arxiv 2505.23992 v1 pith:RJZCY2IL submitted 2025-05-29 eess.SP eess.IVphysics.optics

Weijian Zhang , Hashan K. Weerasooriya , Stanley Chan This is my paper

classification eess.SPeess.IVphysics.optics

keywords single-photonLiDARdeadtimehigh-fluxregimeautoencoderphotonregistrationPDFtimestampsimulationneuralmappingSPAD

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

This paper claims that the slow, photon-by-photon simulation of dead-time-distorted single-photon LiDAR timestamps can be replaced by a learned mapping that is statistically accurate and far faster. The authors model the registered photon count with a Gaussian whose mean is the arrival energy divided by one plus the expected energy lost per registration, and they train an autoencoder to predict the registration probability density directly from the photon flux function. If the claim holds, large-scale high-flux LiDAR datasets for depth estimation and other imaging tasks can be generated in seconds rather than tens of minutes, while retaining the dead-time distortions that make high-flux simulation hard.

What carries the argument

The carrying mechanism is an autoencoder whose encoder compresses the discretized photon arrival flux $\lambda(t)$ into a 16-dimensional latent code and whose decoder reconstructs the registration PDF $f_r(t)$, with the count estimate tied to the PDF through the expected energy loss $\mathbb{E}[g(t_k)] = \langle f_r, g \rangle$ and the Gaussian estimator $\widehat{M}_r \sim \mathcal{N}(R, R/(1+\mathbb{E}[g(t_k)])^2)$, where $R = NQ/(1+\mathbb{E}[g(t_k)])$. This inner-product coupling means the count prediction depends on the accuracy of the predicted registration PDF, so the PDF and count problems are solved as one learned system.

What would settle it

Take environmental parameters not used in training, run the conventional photon-by-photon simulator for many more realizations, say 1000 or more, to obtain a high-confidence reference registration PDF and count distribution, then generate timestamps with the proposed simulator and compare both histograms with a distributional test; if they disagree beyond sampling error in the high-flux regime, the learned mapping is not statistically equivalent to the physical process.

Watch

Extended reading notes

Core claim

The central claim is that the mapping from environmental parameters (signal energy, background level, target delay) to dead-time-distorted registration statistics is learnable: a compact autoencoder fed with the discretized photon arrival flux predicts the 1024-bin registration PDF, and a Gaussian count estimator built from an inner product between this PDF and the per-registration energy-loss function predicts the number of registered photons. The resulting simulator draws a count from this Gaussian and then samples timestamps from the predicted PDF, reproducing the statistics of the conventional sequential rejection simulator while avoiding per-photon culling. The reported PDF prediction RMSE is 0.017 on held-out parameters, including hand-picked parameters outside the training range.

Load-bearing premise

The method's accuracy is judged against averaged histograms from only 20 Monte Carlo realizations of the conventional simulator; if those histograms are noisy or biased estimates of the true registration PDF, the network learns that noise and the reported RMSE does not reflect true physical fidelity.

Editorial extensions

If this is right

Timestamps generated by the proposed simulator can substitute for conventional photon-by-photon output in downstream depth-estimation experiments, with the paper reporting visually identical depth maps while cutting runtime from 35.2 minutes to 8.64 seconds for a 120x180 image.
The network generalizes to environmental parameters outside the training range, since hand-picked test values outside the dataset are predicted accurately and the overall held-out PDF RMSE is 0.017.
Because the simulator's runtime stays nearly flat as the number of laser cycles increases, while the conventional simulator's runtime grows rapidly, large-scale image and video simulation becomes practical.
The Gaussian count model explains the empirically observed bell-shaped registration counts with reduced mean and variance, giving a computable way to decide how many photons to generate before sampling timestamps.

Reading between the lines

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

Editorial inference: the same autoencoder structure could be retrained for other detector behaviors, such as paralyzable dead time, multiple dead-time sources, or detector arrays with per-pixel dead-time variations, replacing the analytic Markov-chain construction in those settings as well.
Editorial inference: because the input flux function already encodes the environmental parameters, one could append system parameters such as dead time and repetition period to the latent code, potentially serving multiple hardware configurations with a single network rather than retraining per system.
Editorial inference: the count estimator's reliance on the predicted PDF means PDF errors propagate into count errors, so a sharper stress test than separate PDF and count metrics is a joint evaluation that measures how count error and timestamp distribution error covary across the parameter space.

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

The paper does what it says: it trains an autoencoder to map the discretized photon arrival flux function to the dead-time-distorted registration PDF, and it adds a Gaussian count estimator. That is a sensible and reasonably new application of a standard architecture to a real problem. The visual results look good, and the speedup is real—35 minutes to 8.64 seconds on a 120x180 image is the kind of number that makes people want to use this. The downstream depth-estimation comparison is a nice sanity check. If I worked on high-flux SPL simulation, I would want this in my toolkit.

The soft spots are mostly about validation, not about the core idea. The ground-truth labels for the PDF are averaged histograms from only 20 Monte Carlo realizations per parameter setting. No confidence intervals, no check that 20 runs are enough, no comparison between independent label batches. That matters because the reported RMSE of 0.017 is computed against those same noisy histograms, so it conflates network approximation error with Monte Carlo label noise. The stress-test note is right: a label-noise-floor experiment is needed before the accuracy claim is trustworthy. Separately, the count estimator in Eq. (6) contains a post hoc variance correction with no derivation. It might be a fine heuristic, but as written it is a fitted knob, not a derived result. That should be flagged clearly. There is also no code or data release, and no runtime comparison against the Rapp et al. Markov-chain PDF method, which is the natural baseline. The parameter ranges are narrow (S, B in [0,3]), so generalization outside that range is unshown.

None of this is fatal. The paper is honest about its assumptions, the architecture is simple enough to reproduce, and the practical benefit is clear. It is a workshop- or short-conference-level contribution with a fixable validation gap. I would send it to peer review, because the idea is useful and the flaws are addressable. My advice to the editor: major revision or conditional accept, with the bar being a label-noise-floor analysis, a derivation or empirical justification for the variance correction, and ideally code release. A serious referee should engage with it rather than desk-reject it.

Referee Report

3 major / 6 minor

Summary. The paper proposes a learning-based fast simulator for single-photon LiDAR registrations under high-flux conditions with detector dead time. The method has two components: a Gaussian model for the registered photon count (Eq. 6) and an autoencoder that maps the discretized photon arrival flux function to the stationary registration PDF fr(t). The autoencoder is trained on averaged histograms from 20 Monte Carlo realizations of a conventional photon-by-photon simulator, with system parameters fixed and environmental parameters {τ,S,B} varied. Experiments report a test RMSE of 0.017 for PDF prediction, visual agreement in Fig. 7, qualitative depth-map similarity in Fig. 1, and a large runtime reduction in Fig. 8.

Significance. If the accuracy and speed claims are properly validated, this work would provide a practical tool for generating large-scale training data for single-photon LiDAR algorithms in the high-flux regime, where conventional simulation is slow. The autoencoder formulation (using the arrival flux as input to resolve the dimensionality mismatch) is a sensible design choice, and the paper is clearly written. The count model, however, contains an empirically tuned variance correction that is not derived, and the entire accuracy evaluation is conducted against noisy 20-realization histograms. These issues currently prevent the central claim of 'statistically accurate' simulation from being fully supported.

major comments (3)

[§4.1, §4.2] The ground-truth labels for both training and evaluation are averaged histograms from only 20 Monte Carlo realizations per parameter setting. The reported RMSE of 0.017 is therefore measured against these noisy labels, and it does not separate the network's approximation error from the Monte Carlo label noise floor. To support the 'high accuracy' claim, the authors should quantify the label noise floor (for example, by computing the RMSE between two independent batches of 20-realization histograms, or by reporting per-bin standard errors) and, for at least a subset of test points, compare predictions against a high-fidelity reference such as a long simulation (e.g., 10,000 realizations) or the Markov-chain PDF from Rapp et al. Without this, the reported accuracy cannot be interpreted as fidelity to the true dead-time registration distribution.
[§3.2, Eq. (6)] The variance in Eq. (6) is introduced as a 'refinement' because the empirical width is narrower than the initial Gaussian approximation, but no derivation is provided. The initial approximation gives variance NQ/(1+E[g])^2 = R/(1+E[g]), while Eq. (6) uses R/(1+E[g])^2. This post hoc adjustment is motivated by the same empirical simulations that are later used to evaluate the count model, making the count accuracy claim partially circular. Please provide a derivation of this variance (e.g., from a renewal-process argument) or validate the refined estimator on an independent dataset not used to tune the correction. The count evaluation should also report quantitative errors (e.g., mean and variance of Mr against long-run simulations) rather than only the visual comparison in Fig. 6.
[§3.3, §4.2] The central claim is that the simulator produces 'statistically accurate' photon timestamps, but the evaluation only checks the marginal PDF fr(t) and the marginal count distribution. The proposed pipeline samples Mr i.i.d. timestamps from the predicted fr(t), whereas the conventional simulator produces timestamps with dead-time-induced dependence (e.g., a minimum gap of td between successive registrations). The paper does not demonstrate that i.i.d. sampling from fr(t) reproduces the joint statistics of the registration process, nor does it quantify whether downstream applications are sensitive to this difference. Please either state explicitly that the simulator targets histogram-based tasks only, or provide a quantitative validation of the generated timestamp statistics (for example, comparing inter-arrival distributions, or reporting depth-estimation RMSE/MAE for the two simulators on the same scenes, extending Fig. 1 beyond visual inspection).

minor comments (6)

[Abstract] The phrase 'the conventional wisdom is computationally intensive' is imprecise; it should be 'the conventional simulation method' or 'the conventional wisdom-based simulation is computationally intensive'.
[§3.2, Eq. (6)] The notation cMr for the estimated count is unusual; please use \hat{M}_r or similar standard notation.
[§4.1] The description of the dataset split and sampling is clear, but please specify whether the input flux function and output PDF are normalized before training, and state the bin width used for the 1024-dimensional discretization.
[§4.2, Fig. 7] The statement that hand-picked θe values 'lie outside the dataset' is ambiguous: since the ranges are continuous, it likely means outside the training set. Please clarify.
[§4.3, Fig. 8] To make the runtime comparison fair and reproducible, please state explicitly whether the autoencoder inference was performed on CPU or GPU (the footnote mentions MATLAB on CPU, but the network was trained on an A100), and if possible report runtime standard deviations over multiple runs.
[Fig. 1] The claim that the depth maps are 'visually the same' would be stronger with quantitative depth errors (root mean square error or mean absolute error) between the two simulators.

Circularity Check

1 steps flagged · score 2.0 of 10

No definitional circularity: the PDF mapping is a supervised surrogate and the count mean has an independent renewal argument, but the count variance is empirically calibrated to the same simulator that provides the evaluation labels.

fitted input called prediction [Section 3.2, Eq. (6); evaluated in Section 4.2]
"However, the empirical width is narrower than q R 1+E[g(tk)]. Therefore, we refine the estimator as cMr ∼ N R, R (1+E[g(tk)])2 ."

The variance of the Gaussian count estimator is not derived from the renewal model; it is adjusted to match the empirical width of the registration-count histograms observed in Section 2.3.1, which are produced by the same conventional simulator that supplies the ground-truth labels for the reported count predictions in Section 4.2. The count 'prediction' is therefore partly a calibration to the target distribution rather than an independent forward prediction. The mean R is still determined through the learned fr and the physical loss function g, so the circularity is partial and does not extend to the PDF mapping itself.

full rationale

The central neural mapping is a standard supervised fit: the autoencoder is trained with MSE against conventional-simulator histograms and evaluated on a held-out 4:1 split, so the reported PDF RMSE of 0.017 is a legitimate internal-consistency check rather than a circular reduction. The count mean in Eq. (6) follows from a genuine renewal-style argument, E[g(tk)]=<fr,g>, with fr supplied by the trained network, so it is not defined in terms of the count labels. Self-citations to the authors' prior work ([8], [13], [23]) are contextual and non-load-bearing; no uniqueness theorem is imported, and no central premise rests solely on the authors' own earlier claims. The only mild circular step is the variance refinement in Eq. (6), where the Gaussian width is calibrated to the empirical count distribution and then compared with the same type of empirical data as 'prediction'. This does not undermine the overall simulator's independent content, but it prevents the count-distribution claim from being fully first-principles. Validation against the conventional simulator rather than real hardware is a correctness/assessment limitation, not a circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central contributions rest on a conventional Poisson-arrival physics model, a stationarity assumption for the registration PDF, and a heuristic Gaussian count estimator whose variance was corrected to match empirical histograms. No new physical entities are introduced.

free parameters (3)

Gaussian count variance correction = variance set to R/(1+E[g(t_k)])^2
Chosen post hoc because empirical Mr width was narrower than the initial CLT variance; no derivation given in Section 3.2.
Autoencoder latent dimension = 16
Architecture choice; no ablations shown, Section 4.1.
Number of Monte Carlo realizations for labels = 20
Ground-truth PDF per sample is the average of 20 conventional simulations; noise level in labels is not quantified, Section 4.1.

assumptions (6)

domain assumption Per-pixel photon arrival is an inhomogeneous Poisson process with flux λ(t)=α·s(t−τ)+λb.
Invoked in Section 2.1, Eq. 1, following prior SPL models; the entire simulation and the AE input depend on this.
domain assumption The SPAD is nonparalyzable with a single dead time source; quantum efficiency is 1 and dark current is 0.
Stated in Section 1 and Section 2.1; the registration PDF and count model assume this hardware model from [11,13].
ad hoc to paper Averaging 20 conventional simulations yields a sufficiently accurate ground-truth PDF for training and evaluation.
Section 4.1 uses 20 realizations per sample as the label; no convergence check or noise bar is provided.
ad hoc to paper The random registration count Mr is approximately Gaussian with variance R/(1+E[g])^2.
Section 3.2, Eq. 6, corrects the CLT variance post hoc; the validity is only checked visually against empirical histograms.
domain assumption The inner product formula E[g(tk)] = ⟨fr, g⟩ correctly gives the expected photon loss per registration.
Section 3.2, Eq. 5; assumes the stationary PDF fr correctly weights the dead-time energy loss.
standard math For large NQ, the Poisson arrival count Ma can be approximated as Gaussian via the central limit theorem.
Used in Section 3.2 to derive the initial Gaussian estimator for Mr.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping." pith.science (2026). https://pith.science/paper/RJZCY2IL

@misc{pith2026250523992,
  author       = {Pith},
  title        = {Pith review of: Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RJZCY2IL}},
  note         = {Machine review of arXiv:2505.23992}
}

read the original abstract

Efficient simulation of photon registrations in single-photon LiDAR (SPL) is essential for applications such as depth estimation under high-flux conditions, where hardware dead time significantly distorts photon measurements. However, the conventional wisdom is computationally intensive due to their inherently sequential, photon-by-photon processing. In this paper, we propose a learning-based framework that accelerates the simulation process by modeling the photon count and directly predicting the photon registration probability density function (PDF) using an autoencoder (AE). Our method achieves high accuracy in estimating both the total number of registered photons and their temporal distribution, while substantially reducing simulation time. Extensive experiments validate the effectiveness and efficiency of our approach, highlighting its potential to enable fast and accurate SPL simulations for data-intensive imaging tasks in the high-flux regime.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages

[1]

Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping

INTRODUCTION Single-photon LiDAR (SPL) is an emerging active imag- ing technique [1, 2], particularly suitable for accurate long- range and low-light tasks because of its high sensitivity and temporal resolution empowered by advanced single-photon avalanche diode (SPAD) sensors and time-correlated single- photon counting (TCSPC) modules [3]. After illumin...

work page Pith review arXiv 2025
[2]

In this section, we start from the photon arrival process and examine factors affecting photon loss and PDF distortions

SPL PHOTON REGISTRATION In a photon timestamp simulator, the most important ingre- dients are the number of photons to generate and their PDF. In this section, we start from the photon arrival process and examine factors affecting photon loss and PDF distortions. 2.1. Photon Arrival from Target For each pixel (i, j), we assume that there is a single bounc...

work page
[3]

METHOD In this section, we formulate the problems and propose our solutions to the challenging mapping tasks of the photon reg- istration number Mr and PDF fr(t), respectively. 3.1. Problem Statement From the previous section, Mr and fr(t) both highly de- pend on a set of parameters θ = {tr, td, σt, N, τ, S, B}. We partition θ into two groups: system para...

work page
[4]

Network and Training Specifications A simple yet effective AE architecture is proposed in Fig

EXPERIMENTAL RESULTS 4.1. Network and Training Specifications A simple yet effective AE architecture is proposed in Fig. 5. The input flux function λ(t) and the output distorted PDF fr(t) are discretized into 1024-dimensional vectors, which can be adjusted based on the actual bin resolution of TCSPC. The encoder progressively downsamples by a factor of 2 ...

work page
[5]

We design a new model for the registration count and learn a neural map- ping from environmental parameters to the registration PDF

CONCLUSION In this work, we introduce a learning-based approach to ac- celerate the simulation of photon registrations in high-flux single-photon LiDAR systems, where detector dead time in- troduces complex distortions to photon statistics. We design a new model for the registration count and learn a neural map- ping from environmental parameters to the r...

work page
[6]

Laser depth measurement based on time- correlated single-photon counting,

J. S. Massa, A. M. Wallace, G. S. Buller, S. Fancey, and A. C. Walker, “Laser depth measurement based on time- correlated single-photon counting,” Optics letters, vol. 22, no. 8, pp. 543–545, 1997

work page 1997
[7]

Advances in single-photon lidar for au- tonomous vehicles: Working principles, challenges, and recent advances,

J. Rapp, J. Tachella, Y . Altmann, S. McLaughlin, and V . K. Goyal, “Advances in single-photon lidar for au- tonomous vehicles: Working principles, challenges, and recent advances,” IEEE Signal Processing Magazine , vol. 37, no. 4, pp. 62–71, 2020

work page 2020
[8]

Single-photon imaging over 200km,

Z.-P. Li, J.-T. Ye, X. Huang, P.-Y . Jiang, Y . Cao, Y . Hong, C. Yu, J. Zhang, Q. Zhang, C.-Z. Peng, F. Xu, and J.-W. Pan, “Single-photon imaging over 200km,” Optica, pp. 344–349, 2021

work page 2021

Show all 29 references

[9]

Rochas, Single Photon Avalanche Diodes in CMOS Technology, EPFL, 2003

A. Rochas, Single Photon Avalanche Diodes in CMOS Technology, EPFL, 2003

2003
[10]

SPAD-Based Sensors,

E. Charbon, M. Fishburn, R. Walker, R. K. Henderson, and C. Niclass, “SPAD-Based Sensors,” in TOF Range- Imaging Cameras, pp. 11–38. Springer, 2013

2013
[11]

The quanta image sensor: Every photon counts,

E. R. Fossum, J. Ma, S. Masoodian, L. Anzagira, and R. Zizza, “The quanta image sensor: Every photon counts,” Sensors, vol. 16, no. 8, pp. 1260, 2016

2016
[12]

Becker, Advanced Time-Correlated Single Pho- ton Counting Techniques, vol

W. Becker, Advanced Time-Correlated Single Pho- ton Counting Techniques, vol. 81 of Springer Series in Chemical Physics, Springer, 2005

2005
[13]

Joint depth and reflectivity estimation using single-photon lidar,

H. K. Weerasooriya, P. Chennuri, W. Zhang, I. Gy- ongy, and S. H. Chan, “Joint depth and reflectivity estimation using single-photon lidar,” arXiv preprint arXiv:2505.13250, 2025

2025 arXiv
[14]

Communication under the poisson regime,

I. Bar-David, “Communication under the poisson regime,” IEEE Transactions on Information Theory, vol. 15, no. 1, pp. 31–37, 1969

1969
[15]

D. L. Snyder and M. I. Miller, Random Point Processes in Time and Space, Springer, 1991

1991
[16]

Dead time compensation for high-flux ranging,

J. Rapp, Y . Ma, R. M. A. Dawson, and V . K. Goyal, “Dead time compensation for high-flux ranging,” IEEE Transactions on Signal Processing, vol. 67, no. 13, pp. 3471–3486, 2019

2019
[17]

High- flux single-photon lidar,

J. Rapp, Y . Ma, R. M. Dawson, and V . K. Goyal, “High- flux single-photon lidar,” Optica, vol. 8, no. 1, pp. 30– 39, 2021

2021
[18]

Parametric modeling and estimation of photon registrations for 3d imaging,

W. Zhang, H. K. Weerasooriya, P. Chennuri, and S. H. Chan, “Parametric modeling and estimation of photon registrations for 3d imaging,” in 2024 IEEE 26th In- ternational Workshop on Multimedia Signal Processing (MMSP), 2024, pp. 1–6

2024
[19]

Mean and variance of single photon counting with deadtime,

D. F. Yu and J. A. Fessler, “Mean and variance of single photon counting with deadtime,” Physics in Medicine & Biology, vol. 45, no. 7, pp. 2043, 2000

2000
[20]

Time correlated single-photon counting (tcspc) using laser excitation,

D. Phillips, R. Drake, D. O’Connor, and R. C. and, “Time correlated single-photon counting (tcspc) using laser excitation,” Instrumentation Science & Technol- ogy, vol. 14, no. 3-4, pp. 267–292, 1985

1985
[21]

Signal processing based pile-up compensation for gated single-photon avalanche diodes,

A. K. Pediredla, A. C. Sankaranarayanan, M. Buttafava, A. Tosi, and A. Veeraraghavan, “Signal processing based pile-up compensation for gated single-photon avalanche diodes,” arXiv preprint arXiv: 1806.07437 , 2018

2018 arXiv
[22]

Sub-picosecond photon-efficient 3d imaging us- ing single-photon sensors,

F. Heide, S. Diamond, D. B. Lindell, and G. Wet- zstein, “Sub-picosecond photon-efficient 3d imaging us- ing single-photon sensors,” Scientific reports, vol. 8, no. 1, pp. 17726, 2018

2018
[23]

Photon- flooded single-photon 3D cameras,

A. Gupta, A. Ingle, A. Velten, and M. Gupta, “Photon- flooded single-photon 3D cameras,” in CVPR, 2019, pp. 6763–6772

2019
[24]

Dead-time cor- rection of fluorescence lifetime measurements and fluo- rescence lifetime imaging,

S. Isbaner, N. Karedla, D. Ruhlandt, S. C. Stein, A. Chizhik, I. Gregor, and J. Enderlein, “Dead-time cor- rection of fluorescence lifetime measurements and fluo- rescence lifetime imaging,” Optics express, vol. 24, no. 9, pp. 9429–9445, 2016

2016
[25]

In- door segmentation and support inference from rgbd im- ages,

N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “In- door segmentation and support inference from rgbd im- ages,” in ECCV, 2012

2012
[26]

Photon-efficient computational 3-d and reflectivity imaging with single-photon detectors,

D. Shin, A. Kirmani, V . K. Goyal, and J. H. Shapiro, “Photon-efficient computational 3-d and reflectivity imaging with single-photon detectors,” IEEE Transac- tions on Computational Imaging, vol. 1, no. 2, pp. 112– 125, 2015

2015
[27]

A few photons among many: Unmixing signal and noise for photon-efficient active imaging,

J. Rapp and V . K. Goyal, “A few photons among many: Unmixing signal and noise for photon-efficient active imaging,” IEEE Transactions on Computational Imag- ing, vol. 3, no. 3, pp. 445–459, 2017

2017
[28]

Resolution limit of single-photon LiDAR,

S. H. Chan, H. K. Weerasooriya, W. Zhang, P. Abshire, I. Gyongy, and R. K. Henderson, “Resolution limit of single-photon LiDAR,” in CVPR, 2024, pp. 25307– 25316

2024
[29]

The role of detection times in reflectivity estimation with single- photon lidar,

R. Kitichotkul, J. Rapp, and V . K. Goyal, “The role of detection times in reflectivity estimation with single- photon lidar,” IEEE Journal of Selected Topics in Quan- tum Electronics, vol. 30, 2024

2024

Pith tools

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

[1] [1]

Ultrafast High-Flux Single-Photon LiDAR Simulator via Neural Mapping

INTRODUCTION Single-photon LiDAR (SPL) is an emerging active imag- ing technique [1, 2], particularly suitable for accurate long- range and low-light tasks because of its high sensitivity and temporal resolution empowered by advanced single-photon avalanche diode (SPAD) sensors and time-correlated single- photon counting (TCSPC) modules [3]. After illumin...

work page Pith review arXiv 2025

[2] [2]

In this section, we start from the photon arrival process and examine factors affecting photon loss and PDF distortions

SPL PHOTON REGISTRATION In a photon timestamp simulator, the most important ingre- dients are the number of photons to generate and their PDF. In this section, we start from the photon arrival process and examine factors affecting photon loss and PDF distortions. 2.1. Photon Arrival from Target For each pixel (i, j), we assume that there is a single bounc...

work page

[3] [3]

METHOD In this section, we formulate the problems and propose our solutions to the challenging mapping tasks of the photon reg- istration number Mr and PDF fr(t), respectively. 3.1. Problem Statement From the previous section, Mr and fr(t) both highly de- pend on a set of parameters θ = {tr, td, σt, N, τ, S, B}. We partition θ into two groups: system para...

work page

[4] [4]

Network and Training Specifications A simple yet effective AE architecture is proposed in Fig

EXPERIMENTAL RESULTS 4.1. Network and Training Specifications A simple yet effective AE architecture is proposed in Fig. 5. The input flux function λ(t) and the output distorted PDF fr(t) are discretized into 1024-dimensional vectors, which can be adjusted based on the actual bin resolution of TCSPC. The encoder progressively downsamples by a factor of 2 ...

work page

[5] [5]

We design a new model for the registration count and learn a neural map- ping from environmental parameters to the registration PDF

CONCLUSION In this work, we introduce a learning-based approach to ac- celerate the simulation of photon registrations in high-flux single-photon LiDAR systems, where detector dead time in- troduces complex distortions to photon statistics. We design a new model for the registration count and learn a neural map- ping from environmental parameters to the r...

work page

[6] [6]

Laser depth measurement based on time- correlated single-photon counting,

J. S. Massa, A. M. Wallace, G. S. Buller, S. Fancey, and A. C. Walker, “Laser depth measurement based on time- correlated single-photon counting,” Optics letters, vol. 22, no. 8, pp. 543–545, 1997

work page 1997

[7] [7]

Advances in single-photon lidar for au- tonomous vehicles: Working principles, challenges, and recent advances,

J. Rapp, J. Tachella, Y . Altmann, S. McLaughlin, and V . K. Goyal, “Advances in single-photon lidar for au- tonomous vehicles: Working principles, challenges, and recent advances,” IEEE Signal Processing Magazine , vol. 37, no. 4, pp. 62–71, 2020

work page 2020

[8] [8]

Single-photon imaging over 200km,

Z.-P. Li, J.-T. Ye, X. Huang, P.-Y . Jiang, Y . Cao, Y . Hong, C. Yu, J. Zhang, Q. Zhang, C.-Z. Peng, F. Xu, and J.-W. Pan, “Single-photon imaging over 200km,” Optica, pp. 344–349, 2021

work page 2021