{"id":"8eaf6dc1-bba9-4f88-801e-fca849084a8e","arxiv_id":"2502.02907","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Silhouette stacking combined with Fourier-domain symmetry detection estimates an object's rotation pole to within about one to three degrees in simulation.","lead":"This paper describes an algorithm that estimates the spin pole of an irregular object, such as an asteroid, by stacking silhouette images taken from different viewing angles. The method could help spacecraft determine a target's rotation during approach without needing high-resolution images or complex 3D models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Robustness claim depends on Assumption III.2, which the paper admits can fail under shadowing; the maximum-symmetry estimate can then be biased.","rationale":"The reader's weakest-assumption analysis identifies Assumption III.2, and my reading agrees: the central claim is that the maximum-symmetry direction in the silhouette-stack amplitude spectrum equals the projected pole, and this equivalence is only guaranteed if the error term has no reflective symmetry of its own. The paper explicitly acknowledges that terminator lines and shadow edges can violate this condition, so the assumption is not merely a technical formality; it is the hinge on which the noise-robustness claim turns. I also note a second, related gap: Theorem III.4 is stated for the expectation of the symmetry score, whereas Algorithm 1 maximizes the score of one observed stack. Even if Assumption III.2 holds on average, the realized argmax can be biased by the same noise fluctuations, and no variance or concentration bound is supplied. This does not invalidate the method—the numerical results provide some empirical support—but it means the theoretical framework is currently weaker than the abstract suggests. The concrete terminator-rotation experiment directly probes the assumption by creating a controlled, realistic shadow geometry independent of the pole direction; if the estimated angle tracks the terminator, the assumption fails. This concern strengthens the reader's CONDITIONAL verdict without moving it to rejection, because the proposed experiment is feasible and the method may still pass it. The paper deserves credit for a simple, parameter-light algorithm and for testing a challenging 90-degree sun phase, but the robustness claim needs this additional check or an explicit caveat about shadow geometry.","tokens_in":27390,"tokens_out":9547,"duration_ms":90129,"concrete_test":"Using the same renderer and the 67P/Bennu models as in Section V, fix image resolution, camera latitude, and the 90-degree sun phase, but vary the sun azimuth (terminator orientation) relative to the pole in 15-degree increments over a full rotation, keeping the true pole-projection angle fixed. Run Algorithm 1 with the Table 1 settings and record the estimated angle for each sun azimuth. If the estimation error varies systematically with the terminator orientation (e.g., peak-to-peak deviation exceeding roughly 2 degrees), Assumption III.2 is violated in a realistic shadowing regime, and the claimed robustness to shadowing is not established. Repeating each configuration with several random centroid-offset realizations would also test whether the single-realization argmax behaves like the expectation in Theorem III.4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption III.2 (Eq. 68) is the load-bearing condition: Theorem III.4's guarantee that the maximum-symmetry angle equals the projected pole requires that the frequency-domain error term A_N^2 has zero expected correlation with its own reflection about every axis. The paper itself concedes this may fail for terminator lines or high-contrast shadow edges (Section III.F.3). Since the estimator is just an argmax of a symmetry score, any systematic symmetry in the error term—e.g., a straight terminator whose image orientation is not aligned with the pole—directly biases the estimated pole-projection angle. The 3-degree error already reported for 67P with perfect alignment at a 90-degree sun phase is consistent with such bias. Additionally, Theorem III.4 only proves the expectation of the symmetry score is maximized at the true pole; Algorithm 1 maximizes the score of a single noisy realization, and no concentration or unbiasedness argument for the argmax is provided. Both gaps bear directly on the central claim of robust pole estimation from maximum symmetry.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PoleStack, a two-stage algorithm for estimating the rotation pole of a principal-axis rotator from silhouette images. Stage 1 stacks silhouettes collected from a hovering camera, computes the DFT amplitude spectrum, and estimates the projected-pole direction as the orientation maximizing a normalized-correlation symmetry score on the log-power spectrum. Stage 2 triangulates several projected-pole measurements from different camera views by solving a linear least-squares problem. The theoretical section derives a reflective-symmetry property of silhouette stacks, models shadowing and registration errors as additive asymmetric terms, and proves, under Assumption III.2, that the expected symmetry score of the noisy amplitude spectrum is maximized at the true projected pole. Experiments use simulated images of Bennu and 67P at a 90-degree sun phase for in-plane estimation, and a Monte Carlo study with synthetic angle noise for 3D triangulation.","tokens_in":27556,"tokens_out":12471,"duration_ms":107508,"significance":"If validated, the method addresses a practical need in small-body approach navigation: an early, landmark-free pole estimate from low-resolution silhouettes that is robust to shadowing and to unknown center-of-mass registration. The paper's strengths include an explicit formalization of the symmetry mechanism, a translation-invariant DFT formulation, and a very simple, precisely specified Algorithm 1. The falsifiable empirical claims, namely degree-level accuracy at a 90-degree sun phase and robustness to centroid alignment, are the right kind of claims for this community. However, the current validation is not yet strong enough for the stated conclusions, and one step of the triangulation derivation needs correction.","major_comments":[{"comment":"The triangulation matrix is derived incorrectly. With the paper's Definition III.6, tan(alpha_j) = (omega^T i_Cj)/(omega^T j_Cj), so each view yields one scalar equation sin(alpha_j)(omega^T j_Cj) - cos(alpha_j)(omega^T i_Cj) = 0, i.e., a single row sin(alpha_j) j_Cj^T - cos(alpha_j) i_Cj^T. The two-row block M_j = [sin(alpha_j) i_Cj^T; -cos(alpha_j) j_Cj^T] in Eq. (92) instead requires, for generic alpha_j, both omega^T i_Cj = 0 and omega^T j_Cj = 0, which forces omega to be parallel to k_Cj and is not equivalent to the angle measurement. Since Section V.B uses Eqs. (87)-(88) as the estimator, the reported Monte Carlo results do not validate the stated triangulation problem; the formulation should be corrected to one row per view and the experiments rerun.","section":"III.G, Theorem III.5, Eqs. (87)-(92)"},{"comment":"The central no-bias guarantee rests on Assumption III.2, which asserts that the squared-amplitude noise A_N^2 is uncorrelated with its own reflection about every axis. The paper itself notes in Section III.F.3 that terminator lines and high-contrast shadow edges can violate this condition. Because Algorithm 1 selects argmax_theta psi(A^2_theta) on a single noisy realization, while Theorem III.4 only proves that the expectation of psi is maximized at the true angle, a shadow pattern with a preferred orientation not aligned with the pole can shift the argmax. No finite-sample concentration bound, adversarial shadow experiment, or empirical check of Assumption III.2 is provided. This gap is load-bearing for the abstract's robustness-to-shadowing claim.","section":"III.F.3, Assumption III.2, Theorem III.4"},{"comment":"The in-plane accuracy claim is supported by only two shape models, a single sun phase of 90 degrees, a single camera latitude of 14 degrees, and one realization per alignment case. The reported 0-3 degree errors are point estimates with no repeated trials, standard deviations, or confidence intervals, so the phrase 'degree-level accuracy' in the abstract and conclusions is not statistically supported. At minimum, the authors should repeat the experiment over independent camera longitude offsets, pole orientations, and sun directions, and report error distributions.","section":"V.A, Figure 8"},{"comment":"The 3D validation injects zero-mean Gaussian noise directly into the projected-pole angles alpha_j and therefore does not exercise the image-based Algorithm 1. Errors produced by the symmetry maximization are not shown to be zero-mean, Gaussian, or independent across views; moreover, the sigma_alpha = 1 degree level used in Figure 11 is below the 3 degree worst-case error observed in Section V.A. An end-to-end test that feeds synthesized image sets from multiple camera views through both stages is needed before the 3D degree-level claim can be accepted.","section":"V.B, Figures 11-13"}],"minor_comments":[{"comment":"In Assumption III.2, the image A^2_theta,N is not defined; the second factor in Eq. (68) appears to be the reflection of A^2_N,theta, but as printed both factors are identical. Please define the reflection operator explicitly.","section":"III.F.3, Eq. (68)"},{"comment":"The substitution phi to tilde-phi in Eqs. (30)-(32) requires that the map from camera longitude to the paired longitude be one-to-one and differentiable; Definition III.12 only asserts existence of a matching longitude. State the additional regularity or give a measure-theoretic proof.","section":"III.B.2, Theorem III.1 proof"},{"comment":"Algorithm 1 applies a circular crop to the amplitude spectrum but does not state whether the spectrum is centered via fftshift before cropping; since the DFT in Eq. (108) is not centered, the implementation should be specified.","section":"IV, Algorithm 1 and Definition VI.8"},{"comment":"There are repeated typographical errors, including 'does notdoesnot' in Definition III.1, 'a the outlier rate' in Section V.B, and 'phi_f (lower)' in the proof of Theorem III.1, which should be corrected in revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the triangulation derivation in Theorem III.5; it is mathematically fixable, but the Monte Carlo claim in Section V.B will need to be regenerated after correction. I would not reject outright, because the symmetry-stacking idea and the DFT formulation are sound and potentially useful to the small-body GNC community. I would ask for a revised version with the corrected triangulation matrix, repeated in-plane trials, and an end-to-end 3D experiment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"PoleStack is a genuinely new, lightweight method for estimating a rotating body's pole from silhouette images alone, and the core insight is right: stack silhouettes from a hovering camera, and the stack is reflectively symmetric about the projected pole; search for that symmetry in the DFT amplitude spectrum and you get translation invariance for free, which sidesteps the unknown-center-of-mass problem that burdens the existing Pole-from-Silhouette method. The algorithm is a few lines, which matters for onboard navigation, and the paper is clearly written.\n\nWhat it does well: the DFT-amplitude trick is the right call, and the related-work positioning is fair. They also flag explicitly where the load-bearing Assumption III.2 (the noise term has no preferred symmetry axis) can fail — terminator lines, shadow edges with a favored orientation — rather than hiding it. The circularity burden is low: the in-plane estimate is a direct symmetry measurement, not a fit to ground truth.\n\nWhere it is soft, in rough order. The validation is the real weakness. Two shapes, a single sun phase (90 deg), a single camera latitude, no repeated trials, no error bars. The worst error (3 deg for 67P, perfect alignment) is a single run. The 3D triangulation study injects synthetic angle noise rather than running the full image pipeline, so the degree-level 3D accuracy claim is inherited from the noise level you assume, not demonstrated end-to-end. There is no baseline comparison to the Pole-from-Silhouette or circle-of-latitude methods that frame the paper, and no released code or data. This is what a referee should push on hardest.\n\nThe theory is weaker than its theorem format suggests, but not fatally. Theorem III.1's proof uses a variable substitution that really requires the longitude interval to be symmetric about the camera longitude; for partial arcs the claim outruns the proof. The empirical partial-arc results (0-1 deg errors at reduced resolution) suggest the conclusion is directionally right anyway, so this is a proof gap, not a broken method. And Theorem III.4 shows the expected symmetry score is maximized at the true pole, while the algorithm maximizes a single noisy realization — no concentration or argmax-bias argument. Assumption III.2 is exactly where a skeptic should point, but I would not read the single 3-deg run as evidence of bias; query-angle discretization and nearest-neighbor rotation artifacts can explain that. The method handled 90-deg sun and centroid misalignment in the runs it did.\n\nMinor: resolving the four-fold ambiguity is handed to external priors. They say so plainly, so it is acceptable, but the standalone method assumes a decent initial guess.\n\nThis is for small-body optical navigation people; a CV generalist can skip it. It deserves a serious referee. The revision bar should be more experiments, a baseline comparison, and either fixing Theorem III.1 for partial arcs or softening its claim.","headline":"New, simple pole-estimation method whose symmetry insight and DFT trick are sound and honestly presented, but the experiments are too thin to fully back the degree-level claims; it deserves a referee, with revision focused on validation and a proof gap.","tokens_in":28081,"tokens_out":5651,"would_cite":true,"duration_ms":147660,"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":"PoleStack estimates the rotation pole of an irregular space object by stacking silhouette images and finding the axis of maximum reflective symmetry in the Fourier amplitude spectrum.","keywords":["pole estimation","silhouette stacking","reflective symmetry","discrete Fourier transform","translation invariance","spacecraft navigation","small bodies","principal-axis rotation"],"falsifier":"Render a synthetic body with a known pole and add a strong, straight shadow terminator whose image-plane orientation is fixed at, say, 30 degrees from the true pole projection but rotated relative to it, then run Algorithm 1 on silhouette stacks with and without that terminator. If the estimated symmetry axis moves toward the terminator orientation by more than the 1-degree query step, the noise model in Assumption III.2 is violated and the maximum-symmetry estimate is biased by structured shadowing.","tokens_in":27159,"feed_emoji":"🛰️","tokens_out":7156,"duration_ms":67590,"temperature":0.7,"pith_summary":"This paper presents PoleStack, an algorithm for estimating the rotation pole of a principal-axis rotator—an object spinning about one of its principal moments of inertia—from silhouette images taken at multiple camera poses. The central assertion is that when silhouettes collected over a range of camera longitudes are co-added into a single image, that stacked image is approximately mirror-symmetric about the direction of the pole projected onto the image plane. The projected-pole direction is therefore recovered by searching for the axis of maximum reflective symmetry in the Discrete Fourier Transform amplitude spectrum of the stack, which preserves the symmetry while making the search invariant to where the object's center appears in the frame. Combining two or more such projected-pole measurements from different camera orientations yields the full 3D pole direction through a simple least-squares triangulation. The paper reports degree-level pole estimates from low-resolution, heavily shadowed, and centroid-aligned synthetic imagery of Bennu and 67P, which matters because early pole knowledge during spacecraft approach enables body-fixed navigation and shape reconstruction.","feed_headline":"PoleStack finds a spinning asteroid's axis from stacked silhouettes","feed_subtitle":"Low-res images, heavy shadowing, and centroid errors still yield degree-level pole estimates.","key_machinery":"The central object is the silhouette-stack image: a co-addition of binary silhouette masks of the rotating body, all expressed in a common camera frame. Two properties of the stack carry the argument. First, its symmetry-preserving component is reflectively symmetric about the projected pole direction (Theorem III.1), so the pole projection is identifiable as the angle of maximum symmetry. Second, the Discrete Fourier Transform amplitude spectrum of the stack preserves that reflective symmetry while being translation invariant (Theorem III.3), so the symmetry search no longer requires knowing the object's center-of-mass pixel location. A low-pass circular crop plus a log compression of the squared amplitude spectrum sharpens the symmetry peak before the per-angle mirror-correlation score is computed.","core_discovery":"The discovery the paper tries to establish is geometric: a silhouette-stack image, defined as the integral or sum of binary silhouette occupancy masks observed across a camera-longitude interval, contains a component that is exactly reflectively symmetric about the projected pole direction, with all deviations from partial longitude coverage, surface shadowing, and registration offsets appearing as an additive error term. Theorem III.1 proves the symmetry for the perfect-silhouette case; Theorem III.2 folds shadowing and registration errors into a symmetric signal plus error; and Theorem III.4 shows that in the squared DFT amplitude spectrum the expected symmetry score is maximized at the true pole projection as long as the error spectrum has no preferred reflection axis. The algorithm then estimates the in-plane pole angle by rotating the log-power amplitude spectrum and scoring each rotation by normalized correlation with its mirror image, and it estimates the full 3D pole by solving a linear least-squares system that triangulates in-plane angles from known camera orientations.","pith_inferences":["The authors leave implicit that the symmetry principle could extend to tumbling objects viewed over short batches, yielding piecewise axis estimates that turn a single approach pass into a continuous pole track.","A direct extension would be to fuse centroid-based frame centering with the translation-invariant amplitude-spectrum scoring, letting residual mis-centering be absorbed in the frequency domain rather than corrected to pixel precision.","Testable extension: the error model predicts that pole bias should grow with the area and coherence of shadow-generated edges in the stack, so a controlled sweep of sun angle and terminator orientation would map the method's domain of validity.","The symmetry search could also serve as an initialization step for shape-from-silhouette or landmark-tracking pipelines, because it produces a body-fixed reference frame earlier than high-resolution feature tracks become available."],"forward_implications":["A single hovering-camera batch is enough to recover the projected-pole direction, with no knowledge of the object's center-of-mass location beyond silhouette extraction.","Two or more batches from different camera latitudes yield the full 3D pole through linear least squares; simulations show degree-level accuracy once camera boresights are separated by about 10 degrees.","The method holds up at a 90-degree sun phase, where self-shadowing heavily corrupts silhouettes, and with brightness-centroid registration, reducing the need for precise attitude or center knowledge.","Reducing image resolution from 1024 by 1024 to 256 by 256 and limiting longitude coverage to a hemisphere leaves pole-projection errors at 0 to 1 degree for the tested bodies.","Adding more camera views lowers variance and outlier rate, with estimates exceeding 5 degrees of error dropping from 1 percent of trials to 0.005 percent as the view count grows from 2 to 4."],"supporting_citations":[{"why":"Defines the Pole-from-Silhouette approach that assigns a computed center of mass and performs grid search; PoleStack's translation-invariant search is positioned against this baseline.","marker":"[8]"},{"why":"Provides the brightness-moment centroid algorithm whose registration errors are one of the two error sources studied.","marker":"[20]"},{"why":"Supplies the FFT algorithm used to compute the Discrete Fourier Transform in the symmetry search.","marker":"[21]"},{"why":"Provides the DFT-based symmetry detection methodology that motivates using amplitude spectra for imperfect symmetries.","marker":"[22]"},{"why":"Supplies a dense optical-flow method suggested for resolving the four-fold pole-direction ambiguity in the amplitude spectrum.","marker":"[25]"},{"why":"Provides the singular value decomposition used to solve the least-squares pole-triangulation system.","marker":"[28]"},{"why":"Supplies the image-rendering tool used to generate the synthetic Bennu and 67P silhouette observations.","marker":"[29]"},{"why":"Supplies the Bennu shape model used as the near-axisymmetric test case.","marker":"[30]"},{"why":"Supplies the 67P/Churyumov-Gerasimenko shape model used as the bilobed, asymmetric test case.","marker":"[31]"}],"fun_headline_variants":["Silhouette stacking reveals asteroid spin axis from low-res images","PoleStack: spin axis from silhouette symmetry despite shadows and noise","Silhouette-stack symmetry yields degree-accurate asteroid spin poles","DFT-amplitude symmetry finds spin poles in shadowed silhouette stacks","PoleStack: degree-level spin axes from stacked silhouettes in shadow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the noise in the silhouette stack not having a preferred mirror axis of its own; if shadows or alignment errors line up along some direction that is not the pole, the maximum-symmetry search can lock onto that wrong direction instead.","fun_headline_variants_meta":{"raw":{"variants":["Silhouette stacking reveals asteroid spin axis from low-res images","PoleStack: spin axis from silhouette symmetry despite shadows and noise","Silhouette-stack symmetry yields degree-accurate asteroid spin poles","DFT-amplitude symmetry finds spin poles in shadowed silhouette stacks","PoleStack: degree-level spin axes from stacked silhouettes in shadow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000632,"raw_usage":{"total_tokens":2888,"prompt_tokens":888,"completion_tokens":2000,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1907}},"tokens_in":504,"tokens_out":2000,"duration_ms":14759,"temperature":1.0,"reasoning_tokens":1907,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:39:26.522825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render a synthetic body with a known pole and add a strong, straight shadow terminator whose image-plane orientation is fixed at, say, 30 degrees from the true pole projection but rotated relative to it, then run Algorithm 1 on silhouette stacks with and without that terminator. If the estimated symmetry axis moves toward the terminator orientation by more than the 1-degree query step, the noise model in Assumption III.2 is violated and the maximum-symmetry estimate is biased by structured shadowing.","supporting_citations":[{"cited_title":"Light-robust pole-from-silhouette algorithm and visual-hull estimation for autonomous optical navigation to an unknown small body,","cited_arxiv_id":null,"evidence_quote":"Defines the Pole-from-Silhouette approach that assigns a computed center of mass and performs grid search; PoleStack's translation-invariant search is positioned against this baseline."},{"cited_title":"Methods of optical navigation,","cited_arxiv_id":null,"evidence_quote":"Provides the brightness-moment centroid algorithm whose registration errors are one of the two error sources studied."},{"cited_title":"O.,The fast Fourier transform and its applications, Prentice-Hall, Inc., 1988","cited_arxiv_id":null,"evidence_quote":"Supplies the FFT algorithm used to compute the Discrete Fourier Transform in the symmetry search."},{"cited_title":"A signal processing approach to symmetry detection,","cited_arxiv_id":null,"evidence_quote":"Provides the DFT-based symmetry detection methodology that motivates using amplitude spectra for imperfect symmetries."},{"cited_title":"Fast optical flow using dense inverse search,","cited_arxiv_id":null,"evidence_quote":"Supplies a dense optical-flow method suggested for resolving the four-fold pole-direction ambiguity in the amplitude spectrum."},{"cited_title":"H., and Van Loan, C","cited_arxiv_id":null,"evidence_quote":"Provides the singular value decomposition used to solve the least-squares pole-triangulation system."},{"cited_title":"Image Rendering and Terrain Generation of Planetary Surfaces Using Source-Available Tools,","cited_arxiv_id":null,"evidence_quote":"Supplies the image-rendering tool used to generate the synthetic Bennu and 67P silhouette observations."},{"cited_title":"The unexpected surface of asteroid (101955) Bennu,","cited_arxiv_id":null,"evidence_quote":"Supplies the Bennu shape model used as the near-axisymmetric test case."},{"cited_title":"The morphological diversity of comet 67P/Churyumov-Gerasimenko,","cited_arxiv_id":null,"evidence_quote":"Supplies the 67P/Churyumov-Gerasimenko shape model used as the bilobed, asymmetric test case."}],"review_version":1}