{"id":"12198799-b613-4cdc-9a98-af3909f80a43","arxiv_id":"2505.02025","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A birotation representation with three axis-aligned basis transformations and manifold optimization is introduced for relative pose estimation, achieving state-of-the-art accuracy on several benchmarks.","lead":"The paper introduces a birotation method for relative pose estimation, representing the motion between two cameras with two rotation matrices and a translation along one of three axes, optimized on the rotation manifold. It reports competitive or better accuracy on standard benchmarks, but the method needs a good initial guess and the code is not yet released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed regularized energy is not actually minimized: update (47) drops the gradient of the L2 term, so the uniqueness argument rests on an objective the algorithm does not solve.","rationale":"The reader's weakest_assumption focuses on iterative descent converging to the correct basin and the selection rule (49) picking the right basis index. My stress-test identifies a more specific and more fundamental problem: the descent step stated in Eq. (47) is not the minimizer of the regularized energy E_i defined in Eq. (36). The missing gradient of the L2 term means the claimed uniqueness mechanism is not operative at convergence, so the method's success depends even more heavily than the reader states on initialization quality and on the unregularized residual landscape. This is an internal inconsistency in the derivation rather than merely a disagreement with outside consensus, and it is directly checkable by re-deriving the normal equations and comparing the two updates numerically. Because the paper's central claim is that minimizing the three energy functions yields an accurate relative pose, this gap is load-bearing. I do not think it warrants rejection, since the empirical results may still hold, but it does warrant conditional acceptance with a request to correct the update derivation or demonstrate numerically that the omission does not change the solutions. The reader's conditional verdict is therefore appropriate, with the concern sharpened to a concrete algebraic issue.","tokens_in":23620,"tokens_out":9291,"duration_ms":108805,"concrete_test":"Re-derive the first-order optimality condition of E_i in Eq. (36) under the left perturbation model used in Section III-D, including the derivative of αR(θ1,i, θ2,i), and obtain the corrected normal equation with the missing αΘ_i term. Implement both the published update (47) and the corrected update on the paper's synthetic datasets from Section IV-F (Gaussian noise and matching error rate sweeps), and also on a small ScanNet subset. If the corrected update produces materially different relative poses or different final energies in any tested regime, the published results are evidence only for a damped unregularized objective, and the uniqueness claim in Section III-D is unsupported. If the two updates converge to the same poses across all tested cases, the omission is benign and the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim depends on the three energy functions E_i in Eq. (36) being minimized on SO(3) and on the L2 regularization term R = ||θ1,i||^2 + ||θ2,i||^2 'ensur[ing] the solution's uniqueness' (Section III-D). However, the normal equation actually solved, Eq. (47), is ΔΘ_i = -(J_i Λ J_i^T + αI)^{-1} J_i Λ e_i. This is the Gauss-Newton step for the unregularized data term e_i^T Λ e_i with αI acting as damping. The gradient of the regularization term with respect to the increment is missing: for the stated objective, the first-order condition includes a term proportional to the current rotation vector (e.g., α(θ1,i, θ2,i)), so the correct equation would have an additional -αΘ_i term on the right-hand side. At any fixed point of the published iteration, J_i Λ e_i = 0, not J_i Λ e_i + αΘ_i = 0. Thus the algorithm does not converge to a stationary point of the regularized energy E_i; α only affects the path, not the limit. This matters because the problem has a one-dimensional gauge freedom (Section III-C) and Section V concedes that 'there remains a potential risk of yielding non-unique solutions in practical applications.' The paper's resolution of that non-uniqueness is attributed to the regularization term, but the regularization term is not actually being minimized in the update. The claim that the three birotation solutions are selected and recovered correctly therefore rests on an unverified and internally inconsistent optimization step, making the convergence-to-the-correct-basin assumption even more load-bearing than the reader's report states.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a 'birotation' parametrization for two-view relative pose estimation. Instead of estimating an essential matrix or a rotation/translation pair directly, the method represents the relative pose through two rotation matrices R1,i and R2,i that rotate the two camera coordinate systems into a configuration related by a pure translation along one of three coordinate axes. Three basis transformations (X-, Y-, and Z-axes) yield three geometric-metric energy functions, which are minimized on SO(3) with a Riemannian iteration; the solution with the lowest weighted residual is selected, and R and t are recovered via R = R2^T R1, t = -s_i R2^T l_i. The paper reports experiments on ScanNet, YFCC100M, KITTI, a pure-rotation dataset, stereo-calibration and visual-odometry tasks, and synthetic robustness tests.","tokens_in":23976,"tokens_out":11721,"duration_ms":88467,"significance":"If the method works as claimed, it provides a genuine alternative to essential-matrix-based two-view geometry that naturally handles pure rotation and near-degenerate configurations; the recovery formulas are derived from the birotation constraints in a self-contained way. The experimental coverage is broad, with dense and sparse correspondences, multiple baselines, ablations, initialization robustness, and synthetic noise studies. The paper's derivation of the residual Jacobians is carefully done, and the plan to release code and datasets is a positive sign. However, the paper's central optimization claim—that the L2 regularization term in the energy is minimized and guarantees uniqueness—is not supported by the stated update rule, and the performance claims need to be delimited because the YFCC100M results do not show universal superiority.","major_comments":[{"comment":"The iteration does not minimize the regularized energy E_i = e_i^T Λ e_i + α(||θ1,i||^2 + ||θ2,i||^2). The normal equation (47) is the damped Gauss-Newton step for the unregularized data term e_i^T Λ e_i only: it is obtained by setting the derivative of the first term of (42) to zero and adding αI as damping. The first-order condition for the stated objective contains an additional term α ∂R/∂Θ_i evaluated at the current iterate, so the published update satisfies J_i Λ e_i = 0 at a fixed point, not J_i Λ e_i + α ∂R/∂Θ_i = 0. As a consequence, α affects the iteration path but not the limit point, and the paper's assertion (Section III-D) that the regularization term 'ensure[s] the solution's uniqueness' is not established. The one-dimensional gauge ambiguity noted in Section III-C and the four-fold ambiguity discussed in Section V are therefore not resolved by the stated algorithm. The authors should either derive and use the correct regularized update (including the gradient of R with respect to the increment under the left-perturbation model) or reframe α as a damping parameter in a Levenberg-Marquardt iteration and substantially weaken the uniqueness claims.","section":"Section III-D, Eqs. (36)-(48)"},{"comment":"The claim of universal superior performance is not supported. On YFCC100M with dense correspondences, the proposed method is below Ling et al. [18] at @5 (45.89 vs 46.59) and @10 (63.45 vs 65.51); with sparse correspondences it is below [18] at all four thresholds (9.49 vs 10.00, 29.19 vs 31.13, 41.67 vs 45.13, 58.22 vs 63.53). The text correctly describes this as 'on par' (dense) and 'second-best' (sparse), but the abstract's 'superior performance' and the Introduction's 'greater accuracy over all other SoTA methods' overstate the results. Please revise the claims to name the regimes in which the method is actually state of the art.","section":"Abstract and Section IV-D1, Table II"},{"comment":"There is an internal tension between the Conclusion's claim that the method 'significantly overcame the limitations of existing birotation-based methods, particularly by ... alleviating the non-uniqueness in relative pose recovery' and the Section V concession that 'there remains a potential risk of yielding non-unique solutions in practical applications.' In light of the first major comment, the regularization term cannot be cited as the mechanism that removes this risk. The authors should either prove a basin-of-attraction or fixed-point uniqueness property for the actual update (47), or explicitly report the frequency of non-unique solutions in the experiments and remove the guarantee language.","section":"Section V and Section VI"}],"minor_comments":[{"comment":"The third contribution says the algorithm 'simultaneously optimizes three energy functions', but Section III-D actually runs three independent optimizations in parallel and then selects among them; please rephrase to 'optimizes three energy functions in parallel'.","section":"Introduction, contributions"},{"comment":"The essential-matrix expression for i=1 has a sign opposite to the derivation from [t]_x R; since the essential matrix is defined only up to scale this is harmless, but the sign should be made consistent or the scale ambiguity stated explicitly.","section":"Section III-D, Eq. (32)"},{"comment":"The citation [36] (Huber) does not appear to describe the 'upper quartile method' used to compute Λ; please provide a precise reference or a description of the outlier-removal procedure.","section":"Section IV-A, paragraph 3"},{"comment":"The manuscript states that source code and datasets will be available upon publication, but no supplementary material or link is included; for review reproducibility, please provide an anonymized code appendix or a working link.","section":"Abstract and Section I"},{"comment":"Fig. 5 is very dense (11 sequences, 3 metrics, 8 methods); consider splitting it into sub-figures or providing the data in table form, since several 'Ours' values are difficult to read.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable extension of the authors' ICRA 2024 birotation calibration work, but the main theoretical selling point—regularization guaranteeing uniqueness—is not actually realized by the algorithm as written, and the empirical claims are overstated on YFCC100M. The experimental effort is commendable, but the optimization section needs a substantive correction or reframing before the paper can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhao et al. extend their earlier stereo-calibration birotation method to general relative pose by adding Y- and Z-axis basis transformations, an arctangent residual that removes pixel-coordinate-dependent weights, and a selection rule over the three solved energies. The geometric derivation of the recovery formula R = R2^T R1, t = -s_i R2^T l_i is clean and self-contained, and the experiments are extensive: ScanNet, YFCC100M, pure rotation, stereo calibration, and KITTI odometry. Pure rotation is a nice win—most essential-matrix methods collapse there, and this one matches homography.\n\nThe soft spot is real and it is in Section III-D. The energy (36) includes the L2 regularizer alpha(||theta1||^2+||theta2||^2), and the text credits it with ensuring uniqueness. But the normal equation actually solved, (47), is the Gauss-Newton step for the data term alone, with alpha I acting as damping. The gradient of the regularizer with respect to the increment is absent—there is no -alpha*Theta term. So at a fixed point of the iteration you get J Lambda e = 0, not J Lambda e + alpha*Theta = 0. The regularizer shapes the path but not the limit. The honest concession in Section V (non-unique solutions are still possible) is consistent with this. This doesn't necessarily sink the method—good initialization and damping can still land in the right basin—but it removes the paper's stated guarantee, and the authors should either revise the update to actually minimize (36) or stop claiming the regularizer resolves the gauge ambiguity.\n\nTwo smaller issues. The abstract's 'superior performance' overstates the YFCC100M dense case, where the method is roughly on par with Ling et al. and worse at @5/@10. And the Zhao [12] global baseline reports near-zero AUC on ScanNet/YFCC—implausibly low for a published QCQP solver, and there is no comment explaining the discrepancy. Code and data are promised 'upon publication,' so independent verification is impossible right now.\n\nWho this is for: anyone working on relative pose or visual odometry who wants a practical alternative to essential-matrix pipelines and doesn't mind a method whose theoretical grounding needs one round of repair. It deserves peer review—the core representation and experiments are worth refereeing—but as a reviewer I would ask for a fix to (47) and a toned-down abstract.","headline":"A genuinely useful three-axis birotation extension with good engineering, but the paper's uniqueness story rests on a regularization term that the published update does not actually minimize.","tokens_in":737,"tokens_out":1852,"would_cite":false,"duration_ms":40314,"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":"The paper claims any relative pose can be recovered by minimizing three energy functions on the rotation manifold SO(3), using a birotation solution that avoids essential-matrix decomposition.","keywords":["relative pose estimation","birotation solution","basis transformations","Riemannian optimization","SO(3) manifold","essential matrix","visual odometry","stereo calibration"],"falsifier":"Compute the three birotation solutions for a large set of image pairs, then restart each energy minimization from many random initial rotations; if the selected solution frequently changes or the relative pose derived from the minimal-energy basis disagrees with the ground truth while another solution agrees, the uniqueness and selection claims would be refuted. A simpler check is a synthetic scene whose true translation is far from all three axes: if at least one energy function systematically fails to converge to the ground-truth pose even from the RANSAC initialization, the coverage claim fails.","tokens_in":23425,"feed_emoji":"📷","tokens_out":7495,"duration_ms":71481,"temperature":0.7,"pith_summary":"The paper sets out to show that relative camera pose, the rotation and translation between two views, can be estimated without ever forming or decomposing an essential matrix. It claims that any such pose can be captured by a birotation model: two rotation matrices that rotate the two camera coordinate systems until the remaining difference is a pure translation along one of three coordinate axes. Three geometric energy functions, one per axis, are minimized on the rotation manifold SO(3), and the lowest-energy solution yields the pose through closed-form recovery formulas. Because the model uses axial basis transformations rather than a single ideal stereo configuration, the paper argues it extends birotation ideas beyond stereo calibration to general indoor, outdoor, pure-rotation, and visual-odometry settings.","feed_headline":"Two rotations recover camera pose better than existing methods","feed_subtitle":"Minimizing three geometric energy functions on the rotation manifold recovers pose without an essential matrix.","key_machinery":"The central object is the $i$-th birotation solution, a tuple $(R_1, R_2, s_i, l_i)$ satisfying $p_1^{C'} = p_2^{C'} + s_i l_i$ after rotating both views, where $l_1, l_2, l_3$ are the X-, Y-, and Z-axis unit vectors. The optimization machinery is Riemannian descent on SO(3): residuals are linearized using left-perturbation Jacobians, the increment $\\Delta\\Theta_i$ is solved in closed form by $\\Delta\\Theta_i = -(J_i \\Lambda J_i^{\\top} + \\alpha I)^{-1} J_i \\Lambda e_i$, and the rotations are updated by exponential maps. The key identity recovers the relative pose from whichever birotation solution has the minimal weighted residual.","core_discovery":"The central claim is that the relative pose $(R,t)$ can be represented as $R = R_2^{\\top} R_1$ and $t = -s_i R_2^{\\top} l_i$, where $R_1$ and $R_2$ are rotations recovered by minimizing $E_i = e_i^{\\top} \\Lambda e_i + \\alpha(\\|\\theta_{1,i}\\|^2 + \\|\\theta_{2,i}\\|^2)$ for $i = 1,2,3$. Here $e_i$ collects arctan-based residuals that measure how far the rotated correspondences are from aligning with a pure translation along the X-, Y-, or Z-axis, $\\Lambda$ is an adaptive outlier mask, and $\\alpha$ regularizes the two rotation vectors. The paper argues that each energy function has a well-defined local minimum on SO(3), that the three candidate minima bracketed by the basis transformations cover general relative poses, and that the residual-weighted selection rule identifies the correct basis. If true, this converts a five-degrees-of-freedom pose problem into a six-degrees-of-freedom birotation problem whose redundant gauge freedom the regularization resolves.","pith_inferences":["If the birotation parametrization is as general as reported, the same three-axis construction could be applied to pose problems beyond two cameras, such as multi-camera rigs or camera-to-robot hand-eye calibration, by treating each pair as a birotation subsystem.","The minimum residual that selects the basis could double as a confidence score: pose estimates with a clearly dominant basis are likely more reliable, which would be useful for outlier rejection or temporal smoothing in SLAM and visual odometry, though the paper does not make this claim.","Because the regularization resolves the redundant sixth degree of freedom, a natural test is whether warm-starting the optimization with the previous frame's solution in a video sequence improves both speed and temporal consistency; the paper does not report such an experiment.","A stronger conjecture is that the arctan-based residuals make the energy landscape benign enough that the RANSAC initialization could be replaced by a small number of random restarts; this is an inference, not a result in the paper."],"forward_implications":["Relative pose can be recovered in closed form from a birotation solution via $R = R_2^{\\top} R_1$ and $t = -s_i R_2^{\\top} l_i$, so no essential-matrix decomposition is needed.","The method handles pure-rotation sequences, where essential-matrix methods degenerate, performing on par with homography-based baselines.","Arctan-based residuals eliminate pixel-location-dependent implicit weighting, improving accuracy for correspondences near image borders and in natural scenes.","With convergence parameters fixed, each iteration costs $O(N)$, and the method outperforms prior birotation and direct pose methods on indoor, outdoor, stereo-calibration, and visual-odometry benchmarks.","The three-energy setup is necessary: ablations show that optimizing only one energy function fails when the true motion is not aligned with that basis axis."],"supporting_citations":[{"why":"Supplies the birotation concept: prior work that first formulated stereo rectification with two independent rotation matrices, which this paper generalizes.","marker":"[3]"},{"why":"Five-point RANSAC solver whose output initializes the optimization and which serves as the primary classical baseline.","marker":"[8]"},{"why":"Global QCQP essential-matrix solver that is the strongest comparison baseline and whose error metrics the paper adopts.","marker":"[12]"},{"why":"Manifold-based essential-matrix estimation baseline used in the comparisons.","marker":"[14]"},{"why":"Direct rotation-translation estimator used as a baseline and as the source of the left-perturbation Jacobian model for the updates.","marker":"[18]"},{"why":"Prior birotation stereo self-calibration method that this work extends; the regularization and uniqueness discussion builds directly on its limitations.","marker":"[19]"},{"why":"Provides the dataset selection protocol and the cumulative pose-error AUC evaluation used in the main comparisons.","marker":"[37]"}],"fun_headline_variants":["Birotation solver ditches essential matrix to pin camera pose","Two rotations on SO(3) recover relative pose more reliably","Relative pose via birotation: energy minima on rotation manifold","New pose solver: three energy maps, two rotations, one answer","Birotation approach rethinks relative pose without essential matrix"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, when started from the five-point RANSAC initialization, the regularized Riemannian descent converges to the correct minimum for at least one of the three energy functions and the selection rule recognizes it; the paper itself concedes in Section V that non-unique solutions remain a potential risk in practice.","fun_headline_variants_meta":{"raw":{"variants":["Birotation solver ditches essential matrix to pin camera pose","Two rotations on SO(3) recover relative pose more reliably","Relative pose via birotation: energy minima on rotation manifold","New pose solver: three energy maps, two rotations, one answer","Birotation approach rethinks relative pose without essential matrix"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3197,"prompt_tokens":951,"completion_tokens":2246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":567,"tokens_out":2246,"duration_ms":16583,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:03:37.962943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the three birotation solutions for a large set of image pairs, then restart each energy minimization from many random initial rotations; if the selected solution frequently changes or the relative pose derived from the minimal-energy basis disagrees with the ground truth while another solution agrees, the uniqueness and selection claims would be refuted. A simpler check is a synthetic scene whose true translation is far from all three axes: if at least one energy function systematically fails to converge to the ground-truth pose even from the RANSAC initialization, the coverage claim fails.","supporting_citations":[{"cited_title":"Online continuous stereo extrinsic parameter estima- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the birotation concept: prior work that first formulated stereo rectification with two independent rotation matrices, which this paper generalizes."},{"cited_title":"An efficient solution to the five-point relative pose prob- lem,","cited_arxiv_id":null,"evidence_quote":"Five-point RANSAC solver whose output initializes the optimization and which serves as the primary classical baseline."},{"cited_title":"An efficient solution to non-minimal case essential matrix estimation,","cited_arxiv_id":null,"evidence_quote":"Global QCQP essential-matrix solver that is the strongest comparison baseline and whose error metrics the paper adopts."},{"cited_title":"Essential matrix estimation using Gauss-Newton iterations on a manifold,","cited_arxiv_id":null,"evidence_quote":"Manifold-based essential-matrix estimation baseline used in the comparisons."},{"cited_title":"High-precision online markerless stereo extrinsic calibration,","cited_arxiv_id":null,"evidence_quote":"Direct rotation-translation estimator used as a baseline and as the source of the left-perturbation Jacobian model for the updates."},{"cited_title":"Dive deeper into rectifying homography for stereo camera online self-calibration,","cited_arxiv_id":null,"evidence_quote":"Prior birotation stereo self-calibration method that this work extends; the regularization and uniqueness discussion builds directly on its limitations."},{"cited_title":"SuperGlue: Learning feature matching with graph neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the dataset selection protocol and the cumulative pose-error AUC evaluation used in the main comparisons."}],"review_version":1}