{"id":"6f164648-e41b-42fd-b037-5aebbe892bdd","arxiv_id":"2501.05446","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"New solvers and a hybrid RANSAC pipeline that estimate relative pose while correcting scale and shift in monocular depth priors, improving pose accuracy for calibrated and uncalibrated cameras.","lead":"This paper shows that explicitly modeling the scale and shift ambiguities in monocular depth predictions yields more accurate relative camera pose from image pairs, beating classic keypoint-only methods on several benchmarks. The result matters because relative pose is a core building block in 3D reconstruction, SLAM, and augmented reality, and the method works with off-the-shelf depth models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The solvers' distance constraints assume one global affine correction per depth map (Eq. 4-5), yet Sec. D concedes shifts vary across regions; no experiment tests the impact of this misspecification on pose accuracy.","rationale":"The reader identified the global affine assumption as the weakest point, and the paper's own Section D confirms the model is an idealization: estimated shifts vary across regions and across different image pairs for the same depth map. This is not a hypothetical concern but an observed property of real depth priors. The central claim that affine corrections improve pose is directly coupled to this model through Eq. (5): the pairwise-distance constraints are exactly valid only under a single per-image affine correction. If the spatial variation of shifts is large, the solver's algebraic core is misspecified and could produce biased poses, undermining the claimed improvements. The paper provides strong empirical support for the benefit of shift modeling (Table 6, Fig. 5, Fig. 6), but it never tests the method under the exact failure mode it concedes. A synthetic experiment that injects controlled spatially-varying shifts into GT depth would settle whether the global affine assumption is load-bearing or merely a convenient approximation. Since the concern is real but not disproven by existing evidence, and the paper already acknowledges the limitation, the conditional verdict remains appropriate. No change to the reader's verdict is needed.","tokens_in":21360,"tokens_out":9989,"duration_ms":101691,"concrete_test":"Using the released code or a reimplementation of the calibrated 3-point solver, generate synthetic two-view scenes with known R, t and GT depth. Add a spatially varying shift to the GT depth, e.g., β(x,y) = c·(x/W)·median(D) (a linear ramp) or per-surface shifts, with c ranging from 0 to 0.5. For noise-free correspondences, run the calibrated solver and a scale-only baseline at each c and measure median pose error over many random scenes. If the affine solver's error rises steeply for c > 0.05–0.1, or becomes worse than scale-only, then the global affine assumption is load-bearing and the claimed improvements may not survive on depth priors with stronger spatial variation. If the error stays near the c=0 level, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that explicitly modeling affine (scale+shift) ambiguities of monocular depth priors improves relative pose—rests on Eq. (4), which posits a single global affine correction per depth map. The pairwise-distance constraints in Eq. (5) are exact only under this model. The authors themselves concede in Sec. D that the fitted shift β can differ across regions/surfaces and even across image pairs for the same depth map, so the global affine model is an idealization. The empirical sections establish that global shifts are large (Fig. 6) and that shift modeling helps (Table 6), but they never quantify the magnitude of the residual spatially-varying component or test the solver's sensitivity to it. If the residual spatial variation is significant, the distance constraints in Eq. (5) become inconsistent: a sampled triple of correspondences spanning different shift regions has no exact solution, and the solver's output (α, β1, β2) is an algebraic compromise that could bias R and t. The method's robustness in this regime is unmeasured, yet it is precisely the regime the authors admit occurs in practice. This is the load-bearing gap: the contribution is defined by the affine model, and the model is known to be violated in a way that is never stress-tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three RANSAC-based solvers for two-view relative pose estimation that jointly estimate the relative pose and per-image affine (scale and shift) corrections of monocular depth priors, covering calibrated, shared-focal, and two-focal camera settings. The solvers are integrated into a hybrid LO-MSAC pipeline that combines depth-aware solvers with classic point-based solvers and both reprojection and Sampson errors. Experiments on ScanNet, MegaDepth, ETH3D, and a generated 2D-3D-S benchmark show consistent improvements over keypoint-only baselines and prior depth-based baselines across multiple matchers and MDE models.","tokens_in":21550,"tokens_out":5437,"duration_ms":58207,"significance":"If the solver derivations are correct, the paper offers a practical and effective way to leverage off-the-shelf monocular depth priors for classical geometric estimation. The empirical study is unusually thorough: multiple datasets, multiple feature matchers, multiple MDE models, ablations of the hybrid components, and runtime measurements are reported, and code is released. The non-circularity burden is also met: pose accuracy is measured against ground-truth pose, and the depth priors come from independently trained models. The main weaknesses are that the solver derivations are asserted rather than demonstrated, and the acknowledged limitation of the global affine model is never stress-tested.","major_comments":[{"comment":"The three solver descriptions omit the actual polynomial systems and only assert solution counts (\"at most 4\", \"8\", \"4\") and template sizes (12x12, 36x36, 40x40). Since the solvers are the central contribution, the manuscript should provide the equations after the substitutions γ=α², ω=1/f², and ω1,ω2, or give a supplementary derivation. In particular, the shared-focal solver \"only select[s] 4 out of the 6 possible equations\" and the two-focal solver \"requires using 5 out of the 6 equations,\" but which equations are selected is not stated. If the choice is arbitrary, this should be justified or shown to be invariant; otherwise readers cannot verify correctness or reproduce the solver.","section":"Sec. 4.1"},{"comment":"The distance constraints in Eq. (5) are exact only under the global affine model in Eq. (4). Section D concedes that the estimated shift β can vary across regions and surfaces and that the same depth map can yield different β groupings for different image pairs. The paper never quantifies the residual spatially varying component after the global fit, nor does it test the solvers' sensitivity to this misspecification. Because Eq. (5) is the foundation of all three solvers, this is a load-bearing gap: a synthetic experiment with per-region shifts, or an analysis of the per-pixel residual after fitting the global affine model, is needed to establish the generality of the claimed improvements and to understand when the distance constraints become inconsistent.","section":"Sec. D and Eq. (5)"},{"comment":"The shared-focal and two-focal solvers are not minimal and drop one or two of the six pairwise distance constraints. The manuscript acknowledges this but does not analyze the consequences: the solved (α, β1, β2) may not satisfy the full set of constraints on the minimal sample, and the subsequent rigid alignment may therefore be biased by the choice of which equations are dropped. The authors state that the hybrid RANSAC pipeline can handle this, but no experiment reports how often minimal samples are inconsistent or whether the dropped-equation choice affects accuracy. At minimum, the paper should specify the equation selection rule and add a synthetic test of solver stability on noise-free data.","section":"Sec. 4.1 and Sec. C"}],"minor_comments":[{"comment":"The notation changes from Eq. (1), with a_i and b_i, to Eq. (4), with α, β1, β2; the relation β_i = b_i/a_i should be stated explicitly in the main text.","section":"Sec. 3"},{"comment":"After the reparameterization γ=α², the paper should state whether α is assumed positive or whether both signs are checked when recovering the scale.","section":"Sec. 4.1"},{"comment":"The axes in Fig. 6 are not labeled with units; the reader must infer that the horizontal axis is the shift normalized by the median depth. Please label the axes consistently with Fig. 5.","section":"Fig. 6"},{"comment":"The rows labeled \"Reference entry - DUSt3R\" and \"Reference entry - MASt3R\" are not clearly integrated with the main comparison; it would be clearer to state in the caption that these are reference results, not baselines run under the same RANSAC protocol.","section":"Tabs. 3-5"},{"comment":"The hyperparameters τr, τs, and λs are described as tuned or fixed without a sensitivity analysis; a short table showing performance under reasonable variations of these thresholds would increase confidence in the conclusions.","section":"Sec. 4.4 and Supp. A"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is strong and the code release is valuable, but the paper currently reads more like a workshop-level system paper than a journal-level solver paper: the polynomial solvers are not actually derived, and the central modeling assumption is acknowledged to be violated without any quantification. I would advise the editor that the revision must add a supplement with the full polynomial formulation and solution-count derivation, plus at least one experiment that probes the global-affine assumption, before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful methods paper and the central claim holds up. The new thing is explicit per-image affine correction (scale plus shift) of monocular depth priors inside relative pose solvers; prior depth-aided solvers such as [4] and [17] modeled scale only. That one change buys consistent, sometimes large, improvements in calibrated, shared-focal, and two-focal settings, indoor and outdoor, with sparse and dense matchers, and — the most interesting empirical finding — it also helps “metric” depth models.\n\nCredit where due. The distance-constraint derivation is clean, the ablation in Table 6 is honest (shift modeling matters most for non-metric priors but still helps metric ones), and the synthetic shift experiment in Fig. 5 is a good way to locate where the benefit kicks in. The hybrid LO-MSAC ablation shows each component earns its place. The limitations section (Sec. D) is candid: the authors concede the global affine model is an idealization, that β varies across regions and surfaces, and that the same depth map can yield different β groupings with different image pairs. Code is promised and runtimes are reported.\n\nSoft spots, in proportion. The stress-test concern is real but overstated as load-bearing. The empirical case does not depend on the affine model being exactly right; it depends on affine correction being a better approximation than scale-only, and the experiments support that consistently. What is genuinely missing is a robustness test: nobody quantifies how the solvers degrade as spatially-varying shift residuals grow, which is precisely the regime the authors admit occurs. Fair ask, not a fatal flaw.\n\nLesser issues: the polynomial solution counts are asserted without derivation; that is normal for solver papers, but a reviewer should ask for the code or a verification. The shared-focal and two-focal solvers drop equations and are not minimal, which the authors acknowledge. And the RANSAC thresholds are tuned per dataset on the test data, which modestly inflates the reported numbers; baselines get the same treatment, so the comparison is still fair in spirit.\n\nThis paper is for anyone doing pose estimation, SLAM, or SfM with learned depth priors. It deserves a serious referee: the math is sound, the experiments are broad, and the limitations are acknowledged rather than buried.","headline":"A solid, well-executed solver paper: explicit affine (scale+shift) correction of depth priors consistently helps relative pose, and the authors honestly flag where their global-affine model breaks down.","tokens_in":22147,"tokens_out":3552,"would_cite":true,"duration_ms":31553,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Estimating the unknown scale and shift of monocular depth maps jointly with camera motion turns depth priors into a reliable geometric cue that consistently improves relative pose estimation over keypoint-only baselines, in both…","keywords":["relative pose estimation","monocular depth priors","affine correction","minimal solvers","hybrid RANSAC","focal length estimation","depth-guided geometry","essential matrix"],"falsifier":"Take image pairs with ground-truth depth and corrupt each depth prior with two different shifts applied to two spatial halves of the scene, sweeping the shift difference from zero up to roughly half the median depth. If the global-affine solver's pose error rises toward or above the point-based baseline as the regional shift difference grows, the single-global-correction model is the limiting factor — the regime the paper's own limitation section says occurs in practice.","tokens_in":21117,"feed_emoji":"📷","tokens_out":12049,"duration_ms":102297,"temperature":0.7,"pith_summary":"Monocular depth models usually predict depth only up to an unknown scale and shift, yet existing geometric pipelines feed those predictions to solvers that assume the ambiguity is just a single scale. This paper argues that ignoring the per-image shift is the main reason depth priors have not helped relative pose estimation, and it develops three solvers that jointly estimate the affine corrections $\\beta_1, \\beta_2$ and the scale ratio $\\alpha$ together with the relative pose — for calibrated cameras, for a shared unknown focal length, and for two unknown focal lengths. The solvers are embedded in a hybrid RANSAC pipeline that also uses classic keypoint solvers, so the depth priors contribute when they are reliable and are safely ignored when they are not. Across indoor and outdoor benchmarks, the hybrid estimator beats keypoint-only baselines and prior scale-only depth-based methods, and the improvement holds across different feature matchers and depth models — including, surprisingly, models trained to predict metric depth.","feed_headline":"Fix depth's hidden scale and shift to sharpen camera pose","feed_subtitle":"Modeling the scale and shift of monocular depth improves relative pose estimation on every tested dataset.","key_machinery":"The load-bearing identity is the pairwise-distance constraint $||\\delta^{(1)}_{jk}||^2 = ||\\delta^{(2)}_{jk}||^2$ between corresponding lifted 3D points: since a rigid transform preserves distances, these equations eliminate rotation and translation and can be solved first for the depth scale $\\alpha$, the shifts $\\beta_1, \\beta_2$, and the focal lengths. The paper re-parameterizes with $\\gamma = \\alpha^2$ and $\\omega = 1/f^2$ to lower the polynomial degree and constructs Gröbner-basis solvers via syzygy-based elimination, yielding at most 4 solutions for the calibrated and two-focal cases and 8 for the shared-focal case. Rotation and translation are then recovered by an orthogonal Procrustes alignment of the affinely corrected 3D points, and the solvers run inside a hybrid LO-MSAC that alternates between depth-aware and classic minimal solvers and jointly minimizes reprojection-plus-Sampson cost.","core_discovery":"The paper's central object is the affine depth-correction model $\\widehat{D}_1 = D_1 + \\beta_1$, $\\widehat{D}_2 = \\alpha(D_2 + \\beta_2)$, in which each monocular depth map is corrected by its own shift and the two maps are related by a single scale ratio before back-projection. Because rigid motion preserves distances, the squared distances between corresponding lifted 3D points must agree across the two views, and these constraints contain no rotation or translation; they form a polynomial system in the affine parameters and any unknown focal lengths. From that system the paper builds three solvers — a minimal calibrated 3-point solver, a shared-focal 4-point solver, and a two-focal 4-point solver — recovers rotation and translation by rigid alignment of the corrected 3D points, and wraps the whole thing in a hybrid locally optimized RANSAC that scores both depth-induced reprojection error and classic Sampson error. On ScanNet, MegaDepth, ETH3D, and a generated two-focal benchmark, the method consistently outperforms keypoint-only baselines and prior scale-only depth solvers, and the shift term proves beneficial even for metric depth priors.","pith_inferences":["A fitted shift value can be read as a per-image bias estimate for a depth model, which suggests a practical diagnostic: large fitted shifts on a 'metric' model flag systematic bias, and an affine fit against ground truth could serve as a model-quality measure for geometry tasks instead of raw metric error.","The same global-affine correction idea should transfer to any task that aligns per-view depth or shape priors, such as point-cloud registration, SLAM front-ends, or multi-view bundle adjustment over the affine parameters, because those pipelines face the identical scale-and-shift ambiguity.","The most natural extension is region-based affine correction — estimating different shifts for different surfaces, guided by segmentation — which would keep the pairwise-distance machinery unchanged while removing the paper's main limitation; the paper itself flags this direction."],"forward_implications":["Modeling the shift, not just the scale, of monocular depth priors is what unlocks their value: the scale-only ablated version loses most of the gain, and the synthetic study shows shifts above 10% of median depth are common even with metric depth models.","The affine-corrected depth cue helps most where keypoint geometry is weakest: image pairs with limited covisibility show the largest pose-error reductions.","The hybrid design makes the gains portable — the same pipeline improves over point-based baselines with SuperPoint+LightGlue, RoMa, and MASt3R matches and with Omnidata, Marigold, Depth-Anything, and MoGe depths, and the gains grow as those models improve.","Uncalibrated settings gain the most: on the two-focal benchmark the median pose and focal errors drop to roughly half the 7-point fundamental matrix baseline, showing that depth priors also anchor focal length estimation.","The pipeline is fast enough for practical use, with median runtimes of 31ms (calibrated), 65ms (shared-focal), and 129ms (two-focal) per image pair on CPU."],"supporting_citations":[{"why":"Supplies the syzygy-based polynomial-solver construction used to build all three depth-aware solvers.","marker":"[34]"},{"why":"The five-point essential-matrix solver, the classic calibrated partner inside the hybrid pipeline and the main baseline to beat.","marker":"[46]"},{"why":"The six-point shared-focal relative pose solver used as the point-based partner and baseline in the shared-focal setting.","marker":"[63]"},{"why":"Provides the seven-point fundamental matrix solver and the Sampson epipolar error used for the fully uncalibrated setting.","marker":"[29]"},{"why":"Prior scale-only relative pose solvers using monocular depth that the paper extends by adding shift modeling.","marker":"[4]"},{"why":"The 3p3d and 4p4d scale-only solvers for fundamental matrix estimation from relative depths, which fail to beat point baselines in the paper's comparisons.","marker":"[17]"},{"why":"The hybrid RANSAC framework (hybrid LO-MSAC) that the paper extends with depth-aware solvers, scoring, and local optimization.","marker":"[12]"},{"why":"The RANSAC implementation that provides the point-based baseline results across all settings.","marker":"[33]"}],"fun_headline_variants":["Affine depth corrections sharpen relative pose estimation","Scale-shift depth fixes outperform keypoint pose solvers","Correcting depth's scale and shift improves pose accuracy","Affine-corrected depth priors enhance relative pose estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Each depth map is assumed to be correct up to a single global scale and a single global shift for the entire image, so every pairwise-distance constraint inherits that one-correction model and would break if the depth error varies across the scene.","fun_headline_variants_meta":{"raw":{"variants":["Affine depth corrections sharpen relative pose estimation","Scale-shift depth fixes outperform keypoint pose solvers","Correcting depth's scale and shift improves pose accuracy","Affine-corrected depth priors enhance relative pose estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000329,"raw_usage":{"total_tokens":1884,"prompt_tokens":1039,"completion_tokens":845,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":781}},"tokens_in":655,"tokens_out":845,"duration_ms":7185,"temperature":1.0,"reasoning_tokens":781,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:12:41.193374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take image pairs with ground-truth depth and corrupt each depth prior with two different shifts applied to two spatial halves of the scene, sweeping the shift difference from zero up to roughly half the median depth. If the global-affine solver's pose error rises toward or above the point-based baseline as the regional shift difference grows, the single-global-correction model is the limiting factor — the regime the paper's own limitation section says occurs in practice.","supporting_citations":[{"cited_title":"Effi- cient solvers for minimal problems by syzygy-based reduc- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the syzygy-based polynomial-solver construction used to build all three depth-aware solvers."},{"cited_title":"An efficient solution to the five-point relative pose problem","cited_arxiv_id":null,"evidence_quote":"The five-point essential-matrix solver, the classic calibrated partner inside the hybrid pipeline and the main baseline to beat."},{"cited_title":"A minimal solution for relative pose with un- known focal length","cited_arxiv_id":null,"evidence_quote":"The six-point shared-focal relative pose solver used as the point-based partner and baseline in the shared-focal setting."},{"cited_title":"Fundamental matrix esti- mation using relative depths","cited_arxiv_id":null,"evidence_quote":"The 3p3d and 4p4d scale-only solvers for fundamental matrix estimation from relative depths, which fail to beat point baselines in the paper's comparisons."},{"cited_title":"Hybrid Camera Pose Estimation","cited_arxiv_id":null,"evidence_quote":"The hybrid RANSAC framework (hybrid LO-MSAC) that the paper extends with depth-aware solvers, scoring, and local optimization."},{"cited_title":"PoseLib - Minimal Solvers for Camera Pose Estimation, 2020","cited_arxiv_id":null,"evidence_quote":"The RANSAC implementation that provides the point-based baseline results across all settings."}],"review_version":1}