{"id":"f9a79204-ccba-43d5-ae7c-ac916e1648ce","arxiv_id":"2412.08135","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A joint extrinsic orientation and gyroscope bias estimator built on normal epipolar constraints can initialize VIO from rotation-only motion before translation parallax arrives.","lead":"This paper presents DOGE, a method that estimates the rotation between camera and IMU together with gyroscope bias during the first moments of visual-inertial odometry, using only rotational motion. It is aimed at AR glasses and drones where physical deformation changes the calibrated rotation, because it removes the usual requirement for translation motion before initialization can finish.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pure-rotation claim lacks an observability condition: for planar pure rotation about a fixed axis, the gyro-bias component along that axis and the extrinsic-orientation component about the same axis are jointly unobservable, so K>=2 can be insufficient.","rationale":"Good faith: the paper is a serious engineering contribution. The simulation and EuRoC experiments consistently show that the rotation-only cost helps, and the failure-detection and weighting details are plausible. The concern is not that the method never works; it is that the strongest claim is broader than the analysis. The planar pure-rotation degeneracy is not a mere proof gap: it is an exact symmetry of the cost. The paper's own counting in Sec. III-A ('NEC can provide three independent constraints... size of K is at least 2') assumes a generic motion, but no genericity condition is stated or tested. The suggested simulation is cheap and directly settles whether the optimization can recover the unobservable components. This is the kind of condition that a revised version should either prove or explicitly exclude, e.g., by requiring at least two non-coaxial rotation axes before the rotation-only stage is declared successful. The existing MODERATE/CONDITIONAL verdict remains appropriate; the paper should be accepted only with this limitation and verification made explicit.","tokens_in":12201,"tokens_out":16869,"duration_ms":188453,"concrete_test":"Simulate noiseless planar pure rotation: keyframe pairs with relative rotations about one fixed camera axis at several speeds; set true R_CI and a true gyro-bias error delta_b collinear with the rotation axis. Evaluate the exact Eq. (12) cost along the two-parameter family (delta_b along n, phi about q) from the true values. If the cost remains exactly zero along this family, the problem is not observable and the 'K at least 2' counting is disproved for this case. Then run the IRLS estimator from multiple initializations on this family; if it cannot move back to the true parameters, the pure-rotation claim needs an excitation condition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the NEC residual of Sec. III-A is a well-posed, informative cost for the six unknowns. It fails in a specific pure-rotation regime. Consider exact pure rotation about a fixed IMU axis n. A relative rotation is A_ij = Exp(theta_ij n); the measured preintegration is \\hat A_ij = A_ij Exp(-delta_b Delta t). In the model the camera rotation is X \\hat A_ij Exp(J_b delta_b) X^{-1}. If delta_b is collinear with n, then Exp(J_b delta_b) commutes with A_ij, so this product equals X A_ij X^{-1} for every delta_b along n. The NEC cost in Eq. (12) depends only on this product; hence it is flat in the bias direction. It is also flat under X -> Exp(phi q) X, because q is the camera rotation axis and this rotation commutes with all A_ij. Thus there are two unobservable directions, the stacked Jacobian or Hessian of the residual cannot have the six independent constraints claimed in Sec. III-A ('size of K is at least 2'), and the minimum of Eq. (12) is a continuum rather than a point. The paper nowhere states the required rotational-excitation condition, so the abstract's 'can be solved in pure-rotation cases' is unsupported for this natural case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DOGE, a VIO initialization method that jointly estimates the camera-IMU extrinsic orientation and gyroscope bias using normal epipolar constraints (NEC). The central idea is to substitute a first-order model of the gyro-preintegrated rotation, parameterized by gyro-bias error δb_g and extrinsic-orientation error δθ_CI, into the NEC cost and minimize the minimum eigenvalue of the moment matrix M_ij. The method adds λ and feature-pair weighting, a Chi-square-based failure detector, and an IESKF/MAP refinement. Experiments on EuRoC and on a simulated pure-rotation scene compare favorably with VINS-Mono and Drt-VIO in accuracy, robustness, and runtime.","tokens_in":12598,"tokens_out":30665,"duration_ms":297327,"significance":"If the rotation-only solvability were established, the method would be a practically useful contribution to VIO initialization under loose or deformed extrinsics, because it would remove the requirement of translation parallax for rotational calibration and would lower initialization latency. The paper is commendable for deriving the coupled constraint, for the explicit weighting/outlier-rejection modules, and for documenting runtimes and failure-detection rates. However, the main theoretical claim is not supported as stated: the proposed cost has a symmetry (and hence a continuum of minima) when the motion is a pure rotation about a fixed axis, so the abstract's assertion that the problem 'can be solved in pure-rotation cases' needs qualification, and the two-keyframe-pair sufficiency argument is incomplete.","major_comments":[{"comment":"The rotation-only cost is not well posed for the pure-rotation cases claimed in the Abstract. Suppose all relative rotations are pure rotations about a fixed IMU axis n. Then every measured preintegrated rotation gamma_hat_ij and every correction Exp(J_b delta_b) (for delta_b parallel to n) is a rotation about n, so the predicted camera rotation X gamma_hat_ij Exp(J_b delta_b) X^{-1} is a rotation about the single camera axis q = R_CI n. Replacing X by X Exp(alpha n) leaves each predicted camera rotation exactly unchanged, because Exp(alpha n) commutes with gamma_hat_ij and with Exp(J_b delta_b). Hence the objective in Eq. (12) is constant along a one-parameter (and, if delta_b is also aligned with n, two-parameter) family of parameter values, lambda_min(M_ij) has a continuum of global minima, and the stacked Jacobian is rank-deficient. This is not an artifact of the first-order approximation in Eq. (9): it is an exact symmetry of the nonlinear cost. The paper should state an explicit rotational-excitation condition (for example, that the set of inter-frame rotation axes spans at least two dimensions) and should qualify the claim that the problem 'can be solved in pure-rotation cases.'","section":"Section III-A, Eqs. (8)-(12)"},{"comment":"The sufficiency argument for K >= 2 is incomplete. Even if the NEC provides three independent constraints per two-view setup for the rotation itself, the constraints in Eq. (12) are functions of the composite quantity X gamma_hat_ij Exp(J_b delta_b) X^{-1}; dependencies among these constraints as functions of the six unknowns are not ruled out. The fixed-axis example above shows that the stacked residual can have rank at most five (or lower) despite K >= 2. Please provide an observability/identifiability analysis of the stacked cost, or a numerical rank test of the Jacobian of Eq. (12), and state the conditions under which a unique global minimum exists.","section":"Section III-A, 'the size of K is at least 2'"},{"comment":"The Jacobian J_theta_CI in Eq. (10) appears to be derived for a left perturbation of R_CI, whereas the manifold operators in Eq. (1) and the state update in Eq. (20) adopt the right-perturbation convention R ⊞ θ = R Exp(θ). Under the right-perturbation R_CI = bR_CI Exp(delta_theta_CI), the derivative of gamma_Ci_Cj with respect to delta_theta_CI is bR_CI (gamma_Ii_Ij^T - I), not bR_CI gamma_Ii_Ij^T bR_CI^T - I. Please clarify the perturbation convention used in the LM optimization and correct Eq. (10) if the right-perturbation convention is intended.","section":"Section III-A, Eq. (10)"},{"comment":"The simulation used to support the pure-rotation claim employs pure rotation with jointly varying roll, pitch, and yaw (Fig. 1), which provides the rotational excitation needed to avoid the fixed-axis degeneracy. It therefore does not validate the unqualified statement that the method works in pure-rotation cases. An additional experiment with rotation about a single fixed axis (e.g., a turntable motion) is needed; with the current formulation, the estimated extrinsic orientation should be non-identifiable in that scenario, confirming the need for an excitation condition.","section":"Section IV-B, Figs. 1 and 6"}],"minor_comments":[{"comment":"'An Extrinsic Orientation and Gyroscope Bias Estimation' is ungrammatical; consider 'Extrinsic Orientation and Gyroscope Bias Estimation' or 'A Method for Estimating Extrinsic Orientation and Gyroscope Bias.'","section":"Title and Abstract"},{"comment":"The sentence introducing Eq. (9) is missing punctuation and the word 'as'; please revise for clarity.","section":"Section III-A, before Eq. (9)"},{"comment":"The IMU preintegration covariance term carries a stray superscript k; the covariance of gamma_Ii_Ij does not depend on the point index.","section":"Eq. (14)"},{"comment":"'cause' should be 'because'; the caption of Fig. 2 also ends without a period.","section":"Section IV-A"},{"comment":"The sequence list 'V1 02, V1 03, V2 02, and V1 03' repeats V1 03; the last item is probably V2 03.","section":"Section IV-B"},{"comment":"The sentence 'bγ_Ci_Cj represents the rotation estimation from frame Ci to frame Cj, which is also equal to bR_CiCj' introduces bR_CiCj before it is defined.","section":"Section III-A"},{"comment":"'Ablation expriment' should be 'Ablation experiment.'","section":"Fig. 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The main technical weakness is the unobservability of the extrinsic orientation about a fixed rotation axis in pure-rotation motion. This is a fixable issue if the authors add an excitation condition and re-scope the claims; I therefore recommend major revision rather than rejection. The paper is otherwise competently written and the experimental comparison is fairly extensive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe new thing in DOGE is a rotation-only estimator that ties extrinsic rotation and gyro bias together through the normal epipolar constraint. That is a real step beyond Drt, which decouples rotation and translation but assumes the extrinsic is known, and beyond older pre-estimating methods that need translation parallax. The paper has solid engineering: the per-feature weighting with covariance propagation and Chi-square gating is well-thought-out, the failure detection keeps undetected bad cases under 1% in their EuRoC test, and the ablation supports the weighting modules. The simulation with pure-rotation segments is a nice demonstration, and the runtimes are competitive.\n\nThe soft spot is theoretical and it is real. The authors claim the joint estimator can be solved in pure-rotation cases and that two keyframe pairs give enough constraints. Neither is proven, and the stress-test example shows why the claim is too strong. If the motion is pure rotation about a fixed axis, the gyro bias component along that axis and the extrinsic orientation component about the corresponding camera axis commute with the relative rotations, so the NEC cost is flat along those two directions. The minimum is a continuum, not a point. The paper nowhere states a rotational-excitation condition. The constraint-count argument from Kneip's NEC paper applies to rotation estimation with known gyro measurements, not to this joint problem with bias and extrinsic in the loop. So the abstract's 'can be solved in pure-rotation cases' is unsupported for a natural degenerate case. This is fixable: add an observability analysis, state the excitation requirements, and soften the claim. Also, the first-order linearization at 20 degrees of initial error is hand-waved; the IRLS plus Cauchy weighting may work in practice, but the paper does not analyze when it will diverge. And no code is released, which makes replication harder.\n\nMy overall take: the core formulation and experiments are worth serious referee time. This is a genuine engineering contribution that will be useful to the VIO community, and the empirical evidence is consistent. But the paper currently overclaims, and the theoretical gap should be addressed before acceptance. I would send it to review with a request for an observability condition and code release.","headline":"A genuinely new rotation-only extrinsic/gyro-bias estimator with strong experiments, but the pure-rotation claim lacks an observability condition and overreaches.","tokens_in":13016,"tokens_out":2376,"would_cite":true,"duration_ms":24919,"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":"A rotation-only constraint can estimate the camera-IMU extrinsic rotation and gyroscope bias together during visual-inertial initialization, before translation parallax exists.","keywords":["visual-inertial initialization","extrinsic orientation estimation","gyroscope bias estimation","normal epipolar constraints","rotation-only estimation","maximum a posteriori","iteratively reweighted least squares","failure detection"],"falsifier":"Run the estimator on a synthetic sequence with pure rotation and zero translation; if it returns biased values even with zero pixel noise, or if the cost is flat along some direction, the rotation-only constraint is not identifiable. A cheaper check is to evaluate the observability matrix of the linearized constraint at a pure-rotation trajectory and inspect its null space.","tokens_in":12042,"feed_emoji":"🔧","tokens_out":10753,"duration_ms":89794,"temperature":0.7,"pith_summary":"The paper claims that the camera-IMU extrinsic rotation and the gyroscope bias can be estimated together during visual-inertial initialization using only rotational motion, before translation parallax exists. The method, called DOGE, builds this estimate from a normal epipolar constraint that depends only on rotation, coupling gyroscope preintegration with the visual bearing vectors. The authors argue that this removes the need to wait for translation excitation and avoids the risk of bundle-adjustment-based post-refinement failing when the initial extrinsic guess is poor. In experiments, when the initial extrinsic rotation error exceeds 5 degrees, DOGE produces lower rotation and bias errors than the compared baselines, and the full VIO system starts with a better trajectory estimate.","feed_headline":"Estimates camera-IMU rotation and gyro bias from rotation-only motion","feed_subtitle":"Works in pure-rotation motion, so visual-inertial startup no longer waits for translation parallax.","key_machinery":"The load-bearing object is the normal epipolar constraint (NEC), which replaces the standard epipolar constraint with the coplanarity of epipolar-plane normal vectors, making the two-view cost depend only on rotation through the minimum eigenvalue of the matrix of normal-vector outer products. DOGE writes the rotation between two camera frames as the extrinsic rotation conjugated with the gyroscope preintegration, then linearizes that expression in the gyroscope bias error and the extrinsic orientation error. That linearized rotation is inserted into the NEC cost, and the minimization of the sum of minimum eigenvalues over keyframe pairs is solved with iteratively reweighted least squares, using an eigenvalue-based weight, a per-feature-pair weight derived from propagated pixel and IMU covariance, and a chi-square test to reject outliers. A maximum a posteriori step, implemented with an iterated error-state Kalman filter, carries the estimates across sliding windows before translation parallax arrives.","core_discovery":"The central claim is that the six-dimensional error state consisting of gyroscope bias correction and extrinsic orientation error is identifiable from normal epipolar constraints alone, even in pure-rotation motion. The paper rewrites the inter-frame rotation as a function of the gyroscope measurements and the extrinsic rotation, linearizes the product in the two error terms, and substitutes this expression into the NEC minimum-eigenvalue cost. The resulting cost is minimized by iteratively reweighted least squares with a chi-square outlier test, and a maximum a posteriori estimator then propagates and refines the estimates across sliding windows until translation parallax accumulates. The paper shows that this yields accurate extrinsic rotation and gyroscope bias estimates with fewer keyframes than methods that wait for translation or that post-optimize with visual-inertial bundle adjustment.","pith_inferences":["The paper does not prove observability; an analytic observability condition for the rotation-only cost would settle whether two keyframe pairs are always sufficient, and would show the motion profiles under which the six unknowns separate.","The weighting strategy is not tied to VIO: the same eigenvalue-based cost with per-feature covariance weighting could be applied to other geometric vision problems that minimize an eigenvalue, such as essential-matrix or homography estimation.","The failure-detection pass rate could be reused as a drift trigger: when the chi-square pass rate drops during operation, the system could re-run the rotation-only calibration rather than only using it at startup."],"forward_implications":["If the rotation-only constraint is identifiable, VIO initialization no longer needs to wait for translation excitation, cutting startup latency in applications such as AR and drone navigation.","Joint estimation of gyroscope bias and extrinsic rotation lets a VIO system recalibrate online after temperature or mechanical deformation changes the camera-IMU mounting, without an offline calibration tool.","The chi-square failure detector gives the system a practical signal to wait for more data, so a bad initialization is less likely to be accepted.","The MAP refinement across sliding windows makes the rotation and bias estimates improve continuously before translation arrives, giving the full VIO estimator a better warm start."],"supporting_citations":[{"why":"Introduces the normal epipolar constraint and shows that rotation can be recovered independently of translation; this is the foundation of the rotation-only cost.","marker":"[11]"},{"why":"Directly optimizes frame-to-frame rotation through the minimum-eigenvalue NEC cost, which DOGE extends to include gyroscope bias and extrinsic orientation.","marker":"[12]"},{"why":"Provides the IMU preintegration model and Jacobians used to express the inter-frame rotation as a function of bias and extrinsic orientation.","marker":"[13]"},{"why":"The rotation-translation-decoupled VIO initialization baseline whose framework and velocity/gravity estimation steps DOGE builds on.","marker":"[3]"},{"why":"The VIO system whose post-estimating initialization serves as the main comparison baseline and whose full pipeline is used for the VIO experiments.","marker":"[6]"},{"why":"Supplies the iterated error-state Kalman filter used for the maximum a posteriori refinement across sliding windows.","marker":"[16]"},{"why":"The unscented transform used to propagate pixel observation noise into full-rank 3D bearing-vector covariances for the weighting strategy.","marker":"[14]"}],"fun_headline_variants":["Camera-IMU rotation and gyro bias estimated from pure rotation","VIO startup no longer waits for translation parallax","Pure rotation suffices for camera-IMU calibration","Extrinsic orientation and gyro bias without translation parallax"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the cost built from two keyframe pairs and a first-order linearization pins down all six unknowns (gyroscope bias and extrinsic rotation) even when the camera is not moving sideways; if that cost is flat or ambiguous, the estimator cannot converge.","fun_headline_variants_meta":{"raw":{"variants":["Camera-IMU rotation and gyro bias estimated from pure rotation","VIO startup no longer waits for translation parallax","Pure rotation suffices for camera-IMU calibration","Extrinsic orientation and gyro bias without translation parallax"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000308,"raw_usage":{"total_tokens":1748,"prompt_tokens":920,"completion_tokens":828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":761}},"tokens_in":536,"tokens_out":828,"duration_ms":8041,"temperature":1.0,"reasoning_tokens":761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:11:47.769160+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the estimator on a synthetic sequence with pure rotation and zero translation; if it returns biased values even with zero pixel noise, or if the cost is flat along some direction, the rotation-only constraint is not identifiable. A cheaper check is to evaluate the observability matrix of the linearized constraint at a pure-rotation trajectory and inspect its null space.","supporting_citations":[{"cited_title":"Finding the exact rotation between two images independently of the translation,","cited_arxiv_id":null,"evidence_quote":"Introduces the normal epipolar constraint and shows that rotation can be recovered independently of translation; this is the foundation of the rotation-only cost."},{"cited_title":"Direct optimization of frame-to-frame rotation,","cited_arxiv_id":null,"evidence_quote":"Directly optimizes frame-to-frame rotation through the minimum-eigenvalue NEC cost, which DOGE extends to include gyroscope bias and extrinsic orientation."},{"cited_title":"A rotation-translation-decoupled solution for robust and efficient visual-inertial initialization,","cited_arxiv_id":null,"evidence_quote":"The rotation-translation-decoupled VIO initialization baseline whose framework and velocity/gravity estimation steps DOGE builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The unscented transform used to propagate pixel observation noise into full-rank 3D bearing-vector covariances for the weighting strategy."}],"review_version":1}