{"id":"dd27fc39-34b5-4599-a006-54a75efdfc92","arxiv_id":"2507.17560","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Under a Gaussian PSF simulation with AST-INPE parameters, Gaussian shape fitting (ALG-6) gives the lowest centroiding error but costs about 36x more compute than simple center-of-gravity variants, while thresholding biases dim-star brightness estimates.","lead":"This paper tests six star-tracker centroiding algorithms on simulated star images and finds that a Gaussian shape-fitting routine is the most accurate but roughly 36 times slower than the lightweight algorithms. The result is a parameter-specific engineering comparison for the AST-INPE star tracker, not a new mathematical method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central ranking is only demonstrated for Gaussian PSFs, the same functional form ALG-6 fits; Section 5.1 admits the real PSF can differ, so the practical claim that ALG-6 is best for AST-INPE is not yet established.","rationale":"The reader's weakest assumption is the right one: the Monte Carlo generator uses a Gaussian PSF, and ALG-6 is the only algorithm that fits a Gaussian. This is not neutral, because it gives ALG-6 a matched-model advantage. The paper is transparent about the limitation in Section 5.1 and even admits the best algorithm could change, but stating a limitation is not the same as testing it. The missing ALG-3/4/5 magnitude errors are also a real completeness problem, and the sentence in Section 4 claiming they should equal ALG-2/2/1 is questionable for ALG-3 because its weighting changes the brightness sum. However, this affects the comparison details more than the core mechanism of ALG-6's advantage. No internal contradiction in the centroiding results was identified, and the computational cost claim is strongly supported by Table 1. The right verdict is therefore unchanged: CONDITIONAL, pending a non-Gaussian PSF test and corrected magnitude reporting.","tokens_in":3893,"tokens_out":5220,"duration_ms":55039,"concrete_test":"Repeat the 10,000-run Monte Carlo of Section 4 with a non-Gaussian PSF matched to the same measured width, e.g., a Moffat profile with beta = 2.5 or an Airy pattern with the same sigma_PSF = 0.663 px, including the same background, readout noise, quantization, and saturation. Recompute Figures 3 and 4 for all six algorithms, with ALG-6 still fitting its Gaussian model to the non-Gaussian images. If ALG-6 retains the lowest centroid and magnitude error, the Gaussian-specific objection is settled; if any of ALG-1 to ALG-5 closes or reverses the gap, the paper should state the ranking as conditional on PSF shape.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the Gaussian PSF used to generate the Monte Carlo images (Section 3, sigma_PSF = 0.663 px) is representative of AST-INPE's true PSF. ALG-6 is a least-squares Gaussian fit (Section 2.6), so the simulation gives it a correctly specified model, while ALG-1 to ALG-5 are generic moments or weighted centroids. Under a misspecified PSF, the comparison could reorder: a Gaussian fit is biased by PSF shape error, and thresholded or weighted algorithms can be more robust to extended or asymmetric cores. Section 5.1 explicitly says that the PSF can depart significantly from a Gaussian distribution and that the best performing algorithm could be different, but no non-Gaussian simulation is run. This is an honest limitation rather than an internal contradiction; nevertheless, the central practical conclusion for AST-INPE depends on resolving it. A secondary reporting problem is Section 4's claim that ALG-3/4/5 magnitude errors should be the same as ALG-2/2/1; for ALG-3 the weighted sum used for A (Section 2.3) is not identical to ALG-2's sum, so that assertion is not self-evidently correct and the missing curves should be computed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports Monte Carlo simulations comparing six centroiding algorithms for the AST-INPE star tracker: five lightweight algorithms (simple thresholded CoG, unthresholded CoG, iteratively weighted CoG, intensity weighted centroiding, and thresholded intensity weighted centroiding) and one Gaussian least-squares shape-fitting algorithm (ALG-6). Synthetic 7x7 pixel images are generated with a Gaussian PSF and include shot noise, readout noise, A/D truncation, and saturation. Centroiding and magnitude-estimation errors are measured against known truth over 10,000 runs per magnitude. The principal finding is that ALG-6 achieves the lowest centroiding and magnitude errors but at a computational cost roughly 35 times that of the light-weight algorithms. The paper explicitly acknowledges that the results depend on the Gaussian PSF assumption and that the ranking could change for other PSFs.","tokens_in":4177,"tokens_out":4631,"duration_ms":49569,"significance":"If the results hold, they provide a useful, transparent benchmark for centroiding-algorithm selection in star tracker development and in related fields such as astrometry and adaptive optics. The Monte Carlo pipeline is clearly documented, the parameter choices are derived from an actual test image, and the comparison of six algorithms under controlled noise conditions is valuable. The computational-cost table, despite its confounds, gives a practical ordering. The main limitation is that the generative model is exactly the model fitted by ALG-6, so the ranking is established only for Gaussian PSFs, and the magnitude-error comparison is incomplete for three of the algorithms. These issues are acknowledged but not resolved within the manuscript.","major_comments":[{"comment":"The magnitude-error comparison is incomplete: errors for ALG-3, ALG-4, and ALG-5 are omitted, with the statement that they 'should be the same' as those of ALG-2, ALG-2, and ALG-1, respectively. This assertion is not self-evident, particularly for ALG-3, whose brightness estimate A is computed using the iteratively weighted intensity I''(x,y) rather than the unweighted I'(x,y); the weighted sum need not equal ALG-2's sum. Please compute and plot these errors, or provide a rigorous derivation showing that the magnitude estimates coincide.","section":"Section 4, Figure 4"},{"comment":"The central claim that ALG-6 is the most accurate algorithm is only demonstrated for a generative model that has the same functional form ALG-6 fits: the simulated PSF is Gaussian (Section 3), and ALG-6 fits a Gaussian shape (Section 2.6). This gives ALG-6 a correctly specified model, and the paper itself concedes in Section 5.1 that the real PSF can depart significantly from Gaussian and that the best algorithm could then differ. Because the practical recommendation for AST-INPE depends on the PSF being nearly Gaussian, the manuscript should include at least one non-Gaussian PSF simulation (e.g., a Moffat profile or an aberrated Gaussian) or a quantitative sensitivity analysis of the ranking to PSF shape.","section":"Section 5.1"},{"comment":"The PSF width sigma_PSF = 0.663 px is derived from a single bright star image (magnitude 2.65) and used for all simulated magnitudes and all centroid locations. The paper does not discuss how field-dependent or magnitude-dependent PSF variations could affect the conclusions, even though the threshold used by ALG-1 and ALG-5 depends directly on the background noise and the PSF sampling changes with width. A sensitivity scan over a plausible range of sigma_PSF values, or at least a stated justification for treating it as fixed, would strengthen the relevance of the ranking for AST-INPE.","section":"Section 3"}],"minor_comments":[{"comment":"There are typos: 'obtanining' should be 'obtaining' and 'couse' should be 'course' in the code availability paragraph.","section":"Section 7"},{"comment":"The notice contains 'preprin t' with an unintended space; please fix the typo.","section":"Abstract / Notice"},{"comment":"The IWCoG weighting formula is difficult to read; the exponent '2𝜎𝑠ℎ 2' is ambiguous. Please rewrite it as a standard Gaussian exponent with an explicit denominator, e.g., exp[-( (x-xc)^2 + (y-yc)^2 )/(2 sigma_sh^2)].","section":"Section 2.3"},{"comment":"The Gaussian PSF is sampled on a 10x10 sub-grid for image generation (Section 3) but ALG-6 uses a 4x4 sub-grid for fitting (Section 2.6). Please state explicitly that this discretization mismatch is small and does not bias the comparison.","section":"Sections 2.6 and 3"},{"comment":"The execution-time measurements include the time to generate the synthetic image, which dominates the runtime of ALG-1 to ALG-5. Since the paper notes that the computational cost of these algorithms 'could not be properly evaluated', consider reporting the incremental algorithm overhead by subtracting the dummy time, so that readers can better compare the algorithms' intrinsic costs.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for an instrumentation-oriented journal and the study is straightforwardly executed. The main concerns are the incomplete magnitude-error reporting and the lack of any non-Gaussian PSF robustness test, both of which are fixable within the manuscript's scope. There is no issue of novelty or citation practice; the work is a benchmark comparison rather than a new algorithm. I recommend major revision because the missing magnitude curves directly undermine one of the two headline claims, and the PSF sensitivity is central to the practical recommendation for AST-INPE."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou can skip this paper if you are not working on star tracker centroiding; it is an engineering report, not a major advance. If you are, it is a transparent Monte Carlo comparison of six standard algorithms specifically for INPE's autonomous star tracker. The headline result: a Gaussian least-squares fit (ALG-6) is the most accurate for both centroid and magnitude estimation, but it is roughly 35x slower than the moment-based algorithms. Among the lighter ones, the known trade-offs hold: IWCoG is poor for intermediate magnitudes, thresholding hurts dim stars, and no-threshold CoG is worst for dim stars.\n\nWhat is new is narrow—the parameter-specific simulation and the timing data. The algorithms and their failure modes are already in the cited literature. That does not make the paper worthless; it makes it a targeted engineering study. The Monte Carlo setup is clearly described, and the paper is honest about its main limitation.\n\nThe soft spots are correctable. Most importantly, magnitude errors for ALG-3, ALG-4, and ALG-5 are not computed. The paper asserts they \"should be the same\" as ALG-2 or ALG-1, but for ALG-3 the amplitude estimate comes from a weighted sum, not the unweighted sum ALG-2 uses, so that claim needs verification, not assumption. Second, the entire benchmark uses a Gaussian PSF, which is the functional form ALG-6 fits. The paper admits in Section 5.1 that the real PSF could differ and reorder the ranking, but it does not test any alternative PSF. That makes the practical claim for AST-INPE conditional. Third, the code is only promised to be public later, not actually available; for a simulation-based paper that hurts reproducibility.\n\nThese are not fatal flaws. The results are likely correct for Gaussian PSFs, and the paper's own acknowledgment of the PSF limitation is a sign of good faith. But a referee should ask for the missing magnitude curves, a non-Gaussian sensitivity check, and the code deposit. With those, it becomes a solid reference for the AST-INPE program.\n\nI would send it to peer review rather than desk reject. The contribution is modest, but the methodology is transparent enough to be checked, and the engineering question is real.","headline":"A transparent Monte Carlo comparison of six standard centroiding algorithms for the AST-INPE star tracker, with a Gaussian least-squares fit winning on accuracy at roughly 35x the cost, but the practical claim is only as strong as the Gaussian PSF assumption.","tokens_in":4723,"tokens_out":4286,"would_cite":false,"duration_ms":44077,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Gaussian least-squares fit wins star-centroid accuracy","keywords":["star tracker","centroiding","Gaussian point spread function","Monte Carlo simulation","shape fitting","center of gravity","magnitude estimation","adaptive optics"],"falsifier":"Rerun the Monte Carlo comparison using a measured or otherwise realistic PSF, such as an Airy pattern or a PSF extracted from real star tracker frames, and check whether ALG-6 remains the most accurate centroiding algorithm; a lightweight algorithm winning under a non-Gaussian PSF would refute the paper's central ranking claim.","tokens_in":3666,"feed_emoji":"🛰️","tokens_out":6625,"duration_ms":63956,"temperature":0.7,"pith_summary":"This paper compares six centroiding algorithms for a star tracker using Monte Carlo simulations of 7x7 pixel synthetic star images with a Gaussian point spread function, shot noise, readout noise, A/D truncation, and saturation. Its central claim is that ALG-6, a least-squares Gaussian shape fit, is the most accurate algorithm both for centroid position and for star magnitude estimation, but also the most expensive, with an average runtime about 36 times larger than the lightweight algorithms. The author shows that thresholding in ALG-1 causes large magnitude errors for dim stars, and that the iteratively weighted algorithms ALG-3, ALG-4, and ALG-5 perform poorly for intermediate magnitudes. The result matters because centroiding accuracy directly limits the attitude accuracy of a star tracker, and dim stars are the most common stars in star tracker images.","feed_headline":"Gaussian shape fit wins star-centroid accuracy","feed_subtitle":"Monte Carlo tests show the Gaussian fit also improves dim-star magnitudes, but costs about 36 times more compute.","key_machinery":"The load-bearing object is the simulated Gaussian PSF with width 0.663 pixels, derived from a single real star image, sampled on a 10x10 sub-grid per pixel to generate realistic star images. Against this model, the paper tests six centroid formulas: thresholded center of gravity (ALG-1), unthresholded center of gravity (ALG-2), iteratively weighted center of gravity using a Gaussian window (ALG-3), intensity-weighted centroiding with exponent q=2 (ALG-4 and ALG-5), and nonlinear least-squares Gaussian shape fitting (ALG-6). The Gaussian shape fit is the mechanism that carries the accuracy advantage: instead of summarizing pixel intensities with moments, it fits the known PSF model directly to the image.","core_discovery":"In the simulated conditions, ALG-6 outperformed ALG-1 through ALG-5 in both centroiding error and magnitude estimation error across the tested magnitude range, with the accuracy gap largest for dim stars, where the thresholded algorithms lose signal below the threshold. For stars brighter than magnitude 1.2, saturation of the central pixels degrades every algorithm. The author states this result is conditional on the Gaussian PSF model, and explicitly notes in Section 5.1 that a real PSF that departs significantly from Gaussian could change the ranking, so the claim is best read as: with a Gaussian PSF and the modeled noise, the Gaussian fit is the most accurate choice.","pith_inferences":["If the true PSF of the star tracker is non-Gaussian, a shape fit matched to that PSF could preserve ALG-6's accuracy advantage, but the paper does not test this; the ranking could also shift toward a lightweight algorithm.","A hybrid strategy that runs ALG-6 only on dim stars, where the accuracy gain is largest, and a cheaper algorithm on bright stars could cut the average runtime while retaining most of the benefit.","The table's runtimes include image generation, so the true computational gap between ALG-1 and ALG-5 is not yet measured; a standalone benchmark could change the relative cost ranking of the lightweight algorithms.","Because magnitude errors follow the same ranking as centroid errors, the paper suggests that a single algorithm can serve both position and brightness estimation, which may simplify star identification logic in an autonomous tracker."],"forward_implications":["For a star tracker whose PSF is close to Gaussian, switching from the current thresholded center-of-gravity algorithm to ALG-6 would improve both centroid accuracy and dim-star magnitude estimates.","The magnitude underestimation in ALG-1 for dim stars is caused by thresholding, so reducing the threshold or using a shape fit could remove the bias without changing the rest of the pipeline.","The runtime penalty of ALG-6, about 5.6 ms per star compared with roughly 0.15 ms for the lightweight algorithms, must be absorbed by the onboard computational budget before ALG-6 can be adopted.","The same accuracy-versus-cost ranking is likely to transfer to astrometry and adaptive optics applications that use similar centroiding algorithms, where the best algorithm depends on the tolerated computational cost.","For bright stars above magnitude 1.2, saturation limits accuracy regardless of algorithm, so no centroiding choice can fix that regime."],"supporting_citations":[{"why":"Describes the autonomous star tracker under development and motivates the search for better centroiding algorithms.","marker":"[1]"},{"why":"Supplies the modified moment algorithm that ALG-1 resembles, establishing the thresholded center-of-gravity baseline.","marker":"[2]"},{"why":"Defines the IWCoG and IWC weighting schemes used by ALG-3 and ALG-4.","marker":"[3]"},{"why":"Reports low-light performance of centroiding algorithms, including the poor intermediate-magnitude behavior of IWCoG that this paper reproduces.","marker":"[4]"}],"fun_headline_variants":["Gaussian fit wins star-centroid accuracy, at 36x compute","Gaussian PSF: shape fit beats lightweight centroiders","Star tracker test: Gaussian fit tops centroid accuracy","Dim stars favor Gaussian fit, but cost more compute","For Gaussian PSF, shape fit edges out simple centroiders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The simulations assume the star's point spread function is Gaussian with a width taken from one real star image; if the actual PSF of the star tracker is significantly non-Gaussian, the ranking of algorithms, and in particular ALG-6's accuracy advantage, could change.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian fit wins star-centroid accuracy, at 36x compute","Gaussian PSF: shape fit beats lightweight centroiders","Star tracker test: Gaussian fit tops centroid accuracy","Dim stars favor Gaussian fit, but cost more compute","For Gaussian PSF, shape fit edges out simple centroiders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1192,"prompt_tokens":738,"completion_tokens":454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":354,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":354,"tokens_out":454,"duration_ms":5498,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:46:46.067876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Monte Carlo comparison using a measured or otherwise realistic PSF, such as an Airy pattern or a PSF extracted from real star tracker frames, and check whether ALG-6 remains the most accurate centroiding algorithm; a lightweight algorithm winning under a non-Gaussian PSF would refute the paper's central ranking claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the autonomous star tracker under development and motivates the search for better centroiding algorithms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the modified moment algorithm that ALG-1 resembles, establishing the thresholded center-of-gravity baseline."},{"cited_title":"B.; Prasad, B","cited_arxiv_id":null,"evidence_quote":"Defines the IWCoG and IWC weighting schemes used by ALG-3 and ALG-4."},{"cited_title":"B.; Prasad, B","cited_arxiv_id":null,"evidence_quote":"Reports low-light performance of centroiding algorithms, including the poor intermediate-magnitude behavior of IWCoG that this paper reproduces."}],"review_version":1}