{"id":"fb7ad74f-bd19-406d-9f42-963caf1841fe","arxiv_id":"1908.06372","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Asynchronous timing between a SPAD detector and laser pulses, using deterministic or randomized offsets, reduces photon pileup and improves single-photon 3D depth accuracy by up to an order of magnitude in bright ambient light.","lead":"Single-photon 3D cameras use lasers and ultra-sensitive detectors to measure distances, but sunlight scrambles their readings. This paper shows that by deliberately desynchronizing the detector's timing windows from the laser pulses, the distortion can be averaged away, improving depth accuracy by up to ten times.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The optimality claim for uniform shifting is proven only for a low-SBR, CLT-approximated upper bound on l0 error, not for RMSE; the 'practically optimal' framing therefore goes beyond what Results 1-2 establish, even though the empirical RMSE gains are separately demonstrated.","rationale":"The paper's central claim is the order-of-magnitude depth-accuracy improvement from asynchronous acquisition, and that claim has independent support: Monte Carlo simulations over wide flux ranges (Figs. 4 and 7) and hardware reconstructions (Figs. 9 and 10). I do not see an internal inconsistency that would overturn the method. The soft spot is the theoretical optimality result. Results 1 and 2 are the only formal justification for choosing uniform shifts, and they optimize a surrogate upper bound on l0 error under idealizations; the RMSE quantity in the headline is connected to this result only indirectly. The reader's verdict already identifies this same assumption, and a conditional verdict is appropriate. My pass does not change that verdict: the concern is real and testable, but it does not undermine the empirical finding, so I recommend UNCHANGED relative to the reader.","tokens_in":21680,"tokens_out":12829,"duration_ms":139039,"concrete_test":"Run a Monte Carlo RMSE comparison under the generative model of Eqs. (1)-(3) with the estimator of Eq. (4), using B=1000, Delta=100 ps, td=50 ns, a fixed total time T, and the same L values as the paper's simulations. Sweep SBR in {0.02, 0.1, 0.5, 1.0} with B*Phi_bkg in {0.1, 1, 11}, and compare uniform shifting against random i.i.d. shifts and two deterministic non-uniform shift sequences with the same L, using at least 10^4 Monte Carlo trials per condition. If uniform shifting attains the minimum modulo-B RMSE in every condition, the surrogate-based optimality is empirically validated; if any alternative beats it, the 'practically optimal' claim must be softened to 'high-performance' for that regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Result 1 (Supp. Note 2) is the load-bearing link between the proposed uniform shifting rule and optimality. It establishes that a constant expected denominator sequence minimizes an upper bound on the average probability of depth error, not RMSE, and the bound is derived under ri approximately equal to Phi_bkg (low SBR), a Gaussian CLT approximation for the estimator difference, and a union bound dominated by the largest 1/E[D_i]. The constant-total-denominator argument also uses the low-SBR approximation, and Result 2 assumes L is a multiple of B to obtain an exactly constant denominator; for finite L < B the denominator sequence is only approximately uniform. The paper then uses this result to justify phrases such as 'theoretically optimal method' (Related Work) and the Section 6 title 'Practically Optimal Acquisition,' which are stronger than the theorem supports. The central empirical claim, order-of-magnitude RMSE improvement, does not depend on uniform shifting being exactly optimal and is supported separately by Monte Carlo simulations and hardware demonstrations. The concern is therefore not that the method fails, but that the optimality framing requires qualification if alternative shift sequences achieve lower RMSE in realistic finite-L, moderate-SBR conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes asynchronous acquisition for SPAD-based 3D imaging, in which the SPAD measurement windows are deliberately shifted relative to the laser pulse train, either by a deterministic uniform sequence or by a stochastic photon-driven sequence. The authors develop a Poisson-multinomial histogram formation model, derive a generalized Coates maximum-likelihood estimator for the incident flux waveform, and prove two results intended to show that uniform shifting yields a constant expected denominator sequence and that this minimizes an upper bound on the probability of depth error. They also analyze active-time optimization and flux attenuation, and validate the approach with Monte Carlo simulations and a hardware prototype. The central empirical claim is that asynchronous acquisition improves depth RMSE by up to an order of magnitude compared with synchronous acquisition with extreme attenuation, across a wide range of ambient-flux conditions.","tokens_in":21888,"tokens_out":17810,"duration_ms":174197,"significance":"If the results hold, this is a valuable contribution to SPAD-based LiDAR: it mitigates pileup during acquisition rather than only in post-processing, and the hardware demonstration under strong ambient light (BΦbkg = 11, SBR = 0.02) is compelling. The paper provides self-contained mathematical derivations for the deterministic-shift estimator, closed-form expressions for the optimal active time and attenuation fraction, and extensive Monte Carlo and experimental comparisons against a strong synchronous baseline. The empirical depth-accuracy gains are convincing and are not obtained by fitting parameters to the data. The main weakness is that the theoretical optimality claims are proved for a surrogate l0 upper bound under several approximations, while the paper's framing at times suggests a stronger optimality guarantee; this needs to be corrected and qualified.","major_comments":[{"comment":"The proof of Result 1 replaces the average over i≠τ of the union-bound terms by a double sum over all i and τ. For i=τ the exponent is zero and each diagonal term equals 1/2, so the displayed '≈' adds B/2 to the bound and makes the subsequent statement that the bound is dominated by the largest 1/E[D_i] invalid: the diagonal terms dominate and contain no denominator dependence. Please exclude the diagonal (or justify that its contribution is negligible) and re-derive the optimization of the bound. This is load-bearing because Result 1 is the stated basis for the uniform-shifting optimality claim.","section":"Supplementary Note 2, Eq. (S6)"},{"comment":"Result 1 is carefully qualified in its formal statement as an upper bound on the average probability of depth error, but the interpretive paragraph claims that a constant expected denominator sequence 'will have lower depth error than all other shifting strategies.' That goes beyond what is proved, since the theorem is established only for a surrogate l0 upper bound under the low-SBR approximation r_i≈Φbkg, a central-limit approximation for q̂_i−q̂_τ, and a union bound. In addition, Result 2 is stated without the condition, supplied only in the proof, that L is a multiple of B; for the finite-L, L<B case used in the simulations (e.g., B=1000 with about 25 cycles), the expected denominator is only approximately uniform, with no bound on the approximation error. Please restate the results with their hypotheses and quantify or bound the finite-L nonuniformity before calling uniform shifting theoretically optimal. The empirical RMSE gains do not depend on exact optimality and can stand on their own.","section":"Section 5, Results 1–2 and 'Interpreting Results 1 and 2'"},{"comment":"The paper repeatedly refers to the generalized Coates estimator for photon-driven shifting, while footnote 6 concedes that the Poisson-multinomial model of Section 4 does not apply when the shift sequence is random. Supplementary Note 5 then asserts the same likelihood factorization 'as before' and gives the closed-form estimator in Eq. (S10) without deriving the likelihood from the actual dependent process, in which shifts are a deterministic function of previous photon arrival times. The estimator may be justifiable from the full likelihood of the arrival-time sequence, but that justification needs to be written out; as it stands, the claim that Eq. (S10) is a maximum-likelihood estimator for photon-driven acquisition is not supported. Please either provide the derivation or clearly label the photon-driven estimator as a heuristic that is validated empirically.","section":"Section 6.2 and Supplementary Note 5"}],"minor_comments":[{"comment":"The word 'aquisition' should be 'acquisition.'","section":"Figure 3 caption"},{"comment":"The notation 'j<i in a modulo-B sense' is ambiguous; please define the set J_{l,i} in the main text as in Supplementary Note 1, and use it consistently in Eq. (3).","section":"Section 4, after Eq. (3)"},{"comment":"Even after excluding the diagonal terms, the step 'dominated by the index i with the largest 1/E[D_i]' is informal; please state explicitly why, under the low-SBR model, the off-peak differences (q_i−q_τ) are identical for all i≠τ, so that the minimum denominator is the controlling quantity.","section":"Supplementary Note 2, Eq. (S6)"},{"comment":"Equation (6) maximizes the total expected denominator, which is used as a surrogate for depth accuracy; the text should consistently say 'optimal active time under the l0 surrogate' rather than 'optimal active time,' especially in the opening of Section 6.1.","section":"Section 6.1 and Supplementary Note 4"},{"comment":"The convergence of the empirical shift distribution to uniform is asymptotic; the practical statement that L≤50 cycles is sufficient should be supported by the denominator-bias simulations in Supplementary Figure 2 or by a mixing-time bound.","section":"Supplementary Note 5, Result 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper is sound and the hardware demonstration is impressive. The main concern is that the theoretical optimality narrative is stronger than the proofs support, particularly the flawed diagonal term in Eq. (S6) and the unqualified interpretation of Results 1 and 2. These issues are fixable within the manuscript's scope, but they affect load-bearing claims and should not be allowed through as-is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the empirical core is real, and the paper deserves a serious referee. The asynchronous acquisition idea—desynchronizing SPAD measurement windows from the laser cycles so pileup distortions average out—is well-motivated, and the hardware demos plus Monte Carlo simulations consistently show order-of-magnitude depth RMSE gains over synchronous acquisition with extreme attenuation. The generalized image formation model and the closed-form MLE for arbitrary deterministic offsets are genuine new pieces; the Supplementary Note 1 derivation is clean.\n\nWhat is actually new: uniform shifting and photon-driven shifting as acquisition design principles, the denominator-sequence analysis that connects shift design to estimator variance, and the active-time optimization. The related-work section honestly distinguishes prior gated and free-running SPAD schemes and says what is added.\n\nThe soft spots match the stress-test note, and they are concentrated in the optimality story. Result 1 proves that a constant expected denominator sequence minimizes an upper bound on the l0 depth-error probability under a low-SBR approximation, a Gaussian CLT, and a union bound. That is not RMSE optimality. The Section 6 title 'Practically Optimal Acquisition' and the Related Work phrase 'theoretically optimal method' outrun what the theorem establishes. The photon-driven estimator in Eq. S10 is also not a true MLE in the strict sense: the PMD likelihood of Section 4 assumes deterministic shifts, which the paper acknowledges, so the closed-form Coates-style estimator for photon-driven shifting is a well-tested heuristic, not an MLE. Hardware experiments in Figs. 9-10 are convincing but lack repeated-trial error bars. Both of these are addressable and do not touch the central empirical finding.\n\nThe central claim—asynchronous acquisition reduces depth error by up to an order of magnitude in high ambient light—is supported separately from the optimality result, with both simulation and hardware. The design rules remain useful even if uniform shifting is only near-optimal in practical finite-L regimes. I would cite this paper and would bring it to the reading group.\n\nFor peer review: send it out. Ask the authors to qualify the optimality language, clarify that the photon-driven variant uses a heuristic estimator, and add error bars to the hardware plots. That is a revision, not a rejection.","headline":"The empirical case for asynchronous SPAD 3D imaging is solid and the depth accuracy gains are real; the 'theoretically optimal' framing overstates what Result 1 proves, but that is fixable and does not change the verdict.","tokens_in":22411,"tokens_out":2463,"would_cite":true,"duration_ms":23057,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Deliberately mismatching the SPAD detector window and the laser cycle averages out photon pileup and restores depth accuracy in bright ambient light.","keywords":["asynchronous acquisition","SPAD time-of-flight depth imaging","photon pileup","generalized Coates estimator","uniform shifting","photon-driven shifting","active time optimization","flux attenuation"],"falsifier":"Measure the empirical per-bin denominator $E[D_i]$ under uniform shifting with a number of SPAD cycles not divisible by the number of bins; Result 2's approximation predicts these values are nearly equal, so if a deliberately non-uniform shift sequence with the same total denominator produces lower depth RMSE under the low-background-light, near-Gaussian regime, Result 1's optimality claim would be refuted.","tokens_in":21438,"feed_emoji":"🔦","tokens_out":8886,"duration_ms":84027,"temperature":0.7,"pith_summary":"The paper proposes asynchronous acquisition for SPAD time-of-flight 3D cameras: instead of always starting the detector timing window at the same point in every laser cycle, the window is shifted by deterministic or randomized offsets. The paper's claim is that this spreads the distortion known as photon pileup across all histogram bins and lets the true laser-return peak accumulate coherently, so depth estimates remain accurate even under bright ambient light. It develops a generalized image formation model, a closed-form maximum-likelihood Coates estimator for arbitrary shifts, and theoretical results showing that uniform shifting gives a constant denominator sequence, which minimizes an upper bound on the probability of depth error. Simulations and hardware experiments show up to an order of magnitude lower depth RMSE compared with synchronous acquisition with extreme attenuation.","feed_headline":"Stagger the SPAD window and erase pileup in 3D imaging","feed_subtitle":"Misaligning detector and laser cycles averages out pileup, improving depth accuracy by up to 10x under ambient light.","key_machinery":"The load-bearing object is the denominator sequence $D_i$: the number of SPAD cycles in which histogram bin $i$ is still active, meaning no photon was detected before it. The generalized Coates estimator $q_i = N_i / D_i$ is the maximum-likelihood estimate of the per-bin photon-arrival probability, and pileup is exactly what drives $D_i$ to near zero in late bins. Uniform shifting makes the expected denominator $E[D_i]$ equal across bins, which the paper proves minimizes an upper bound on the probability of depth error and makes error roughly depth-invariant. The two companion mechanisms are the optimal active time $m_{opt}$, which maximizes total expected denominator under a fixed acquisition time, and photon-driven shifting, whose random cycle lengths drive the shift sequence to a uniform stationary distribution.","core_discovery":"The central discovery is that photon pileup in SPAD time-of-flight depth cameras can be largely prevented at acquisition time instead of only corrected after the fact. Deliberately cycling the detector's timing window through offsets that span one full laser period redistributes the exponentially decaying bias caused by early-arriving ambient photons across all histogram bins, while the reflected laser pulse stays at a fixed bin and accumulates coherently. The paper proves that, in the low signal-to-background regime, a uniform shift sequence yields a constant expected denominator sequence (Result 2), which in turn minimizes an upper bound on the probability of depth error (Result 1), and that free-running photon-driven shifting produces the same uniformity asymptotically (Result 3). Simulations and experiments with a fast-gated SPAD show up to an order-of-magnitude lower depth RMSE than synchronous acquisition with extreme attenuation.","pith_inferences":["Because the constant-denominator argument is independent of the true depth, asynchronous acquisition should make depth-error RMSE roughly flat across the full unambiguous range; this is testable by plotting per-depth-bin RMSE, and the paper's supplementary figure already shows this trend.","The same 'shift to average out pileup' recipe should transfer to other TCSPC-based active imaging, for example fluorescence lifetime imaging or non-line-of-sight imaging, although the paper cautions that non-uniform shift sequences may be needed when the incident waveform is an exponential decay or an arbitrary transient rather than a periodic delta pulse.","An online adaptive acquisition controller—estimate the ambient flux with a few early cycles, then switch between uniform shifting, photon-driven shifting, and optimal attenuation—would combine the regimes the paper analyzes separately and is a natural next step.","The photon-driven uniformity result is asymptotic in the number of cycles, so quantifying its finite-time mixing rate would tell practitioners how long acquisition must run before the depth-dependent denominator bias disappears."],"forward_implications":["SPAD-based LiDAR can recover accurate depth in strong ambient light without heavy optical attenuation, because asynchronous shifting redistributes pileup rather than merely correcting it.","The generalized Coates estimator provides a closed-form, non-iterative way to estimate the incident flux waveform from arbitrarily shifted histograms, making real-time depth estimation feasible.","Uniform shifting combined with the optimal SPAD active time gives up to a factor-of-6 RMSE improvement over using the full active window, and up to an order-of-magnitude improvement over synchronous methods overall.","Photon-driven shifting adapts automatically to per-pixel albedo variations, so one acquisition setting works for both bright and dark objects without per-pixel attenuation tuning.","Combining photon-driven shifting with optimal flux attenuation further reduces depth error, in some settings to near zero."],"supporting_citations":[{"why":"Defines the synchronous baseline and the optimal-flux/attenuation model that the paper's asynchronous methods are compared against and extend.","marker":"[13]"},{"why":"Supplies the original Coates pileup correction that the generalized Coates estimator (Eq. 4) generalizes to asynchronous shifts.","marker":"[8]"},{"why":"Provides the Poisson-multinomial distribution used to write the joint histogram likelihood and justify the MLE.","marker":"[10]"},{"why":"Models free-running/dead-time acquisition as a Markov chain, the basis for photon-driven shifting and its asymptotic uniformity result.","marker":"[28]"},{"why":"Describes the fast-gated SPAD detector used in the experimental prototype and enables active-time optimization.","marker":"[6]"},{"why":"Provides the gated-SPAD pileup model and denominator-style correction that the asynchronous estimator generalizes.","marker":"[25]"},{"why":"Presents regularized/iterative pileup-corrected depth estimation cited as a complementary approach and extension path for spatial priors.","marker":"[14]"}],"fun_headline_variants":["Asynchronous SPAD windows average out photon pileup","Cycle detector timing to kill pileup in 3D cameras","Misaligned laser and detector boost depth accuracy 10x","Shift SPAD windows to spread ambient light noise","Randomized offsets erase SPAD pileup bias"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that uniform shifting is the optimal strategy assumes background photons dominate signal photons at every time bin, treats the estimator's error as roughly Gaussian, and uses an upper bound on the chance of choosing the wrong depth bin as a stand-in for actual depth error; if those approximations fail, some other shift pattern could beat uniform shifting.","fun_headline_variants_meta":{"raw":{"variants":["Asynchronous SPAD windows average out photon pileup","Cycle detector timing to kill pileup in 3D cameras","Misaligned laser and detector boost depth accuracy 10x","Shift SPAD windows to spread ambient light noise","Randomized offsets erase SPAD pileup bias"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3207,"prompt_tokens":909,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2220}},"tokens_in":525,"tokens_out":2298,"duration_ms":14386,"temperature":1.0,"reasoning_tokens":2220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:47:22.750987+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the empirical per-bin denominator $E[D_i]$ under uniform shifting with a number of SPAD cycles not divisible by the number of bins; Result 2's approximation predicts these values are nearly equal, so if a deliberately non-uniform shift sequence with the same total denominator produces lower depth RMSE under the low-background-light, near-Gaussian regime, Result 1's optimality claim would be refuted.","supporting_citations":[{"cited_title":"Photon-flooded single-photon 3d cameras","cited_arxiv_id":null,"evidence_quote":"Defines the synchronous baseline and the optimal-flux/attenuation model that the paper's asynchronous methods are compared against and extend."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the original Coates pileup correction that the generalized Coates estimator (Eq. 4) generalizes to asynchronous shifts."},{"cited_title":"On the structure, covering, and learning of pois- son multinomial distributions","cited_arxiv_id":null,"evidence_quote":"Provides the Poisson-multinomial distribution used to write the joint histogram likelihood and justify the MLE."},{"cited_title":"Time-gated single- photon detection module with 110 ps transition time and up to 80 mhz repetition rate","cited_arxiv_id":null,"evidence_quote":"Describes the fast-gated SPAD detector used in the experimental prototype and enables active-time optimization."},{"cited_title":"Sub-picosecond photon-efficient 3d imaging us- ing single-photon sensors","cited_arxiv_id":null,"evidence_quote":"Presents regularized/iterative pileup-corrected depth estimation cited as a complementary approach and extension path for spatial priors."}],"review_version":1}