{"id":"f9b113bb-552b-4f6d-87aa-8e675efe00b8","arxiv_id":"2507.17998","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The authors derive an optimizable registration cost for lines and planes as projection residuals on the affine Grassmannian and validate it on object registration, RGB-D odometry, and PnL problems.","lead":"This paper derives a cost function for aligning 3D lines and planes by measuring distances on the Grassmann manifold, a curved space of subspaces, and uses it in a branch-and-bound solver that is robust to outliers. The cost is written directly in terms of rotation and translation, making it usable in standard optimizers for registration tasks such as camera pose estimation and odometry.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global optimality of the translation BnB rests on an unproven vertex-max bound in App. 8.4.1; a random search over t in the cube can settle whether Eq. (62) actually bounds Eq. (60).","rationale":"I agree with the reader's weakest_assumption: the soft spot is not the core cost derivation. Theorem 1, Theorem 2, and the zero-equivalence in Problem 2 are internally consistent; Eq. (7) is a differentiable, basis-invariant projection cost whose zero set matches zero Grassmann distance. The chordal-versus-geodesic overstatement is real, but I treat it as a presentation issue because the paper's central claim concerns the zero set and the experiments compare relative performance; replacing 'geodesic' with 'projection/chordal' does not invalidate Eq. (7) as an optimizable registration cost. The translation BnB bound is different: it supports a positive guarantee ('deterministically optimal') and is explicitly justified by an empirical observation rather than a theorem. Moreover, as written, the image of the translation cube under the relevant affine map is not a 1D arc, so the vertex-max argument is not even geometrically established. If the proposed random search finds a violation, the solver must be relabeled heuristic or the bound corrected; either way the paper remains conditional rather than rejected, because the cost and the empirical improvements stand independently. Hence the reader's CONDITIONAL verdict is unchanged.","tokens_in":24716,"tokens_out":17175,"duration_ms":203112,"concrete_test":"Run a randomized counterexample search for the line-to-line case: for each trial, sample R*, unit vectors d, b1, b2 with b1 perpendicular to d, a center t0, and cube half-side sigma in {0.05, 0.1, 0.5, 1, 2}; also center trials at random points inside an initial cube so that subcubes are tested. Draw 10^6 t uniformly from C_t and compare ||P_{z((R*,t0))}tilde_b2 - P_{z((R*,t))}tilde_b2|| with the maximum over the 8 vertices (Eq. 62). If any interior t exceeds the vertex maximum, Eq. (62) is false and the BnB lower bound is invalid; if no violation appears across a large random sample, the gap remains unproven and a proof is still required before claiming global optimality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The advertised 'globally optimal' solution depends on the translation lower bound in Appendix 8.4.1. To obtain Eq. (63), the paper needs an upper bound for ||P_{z(T0)} tilde_b2 - P_{z(T)} tilde_b2|| over the whole cube C_t. Eq. (62) asserts that this maximum is attained at a vertex because the inner products (tilde_b2)^T tilde_b1'(R*, t) are only 'monotonic increasing, monotonic decreasing, convex, or concave', justified by 'in every tested case in our experiments' rather than by proof. The geometric statement preceding it is also suspect: t -> R*b1 + R*(I - dd^T)R*^T t maps the 3D cube through a rank-2 linear map, so the image is a parallelogram, and after normalization a 2D surface on S^3, not a line segment with two endpoints. A concave inner-product profile can place the norm's maximum in the interior; if this happens, Eq. (63) overestimates the lower bound and BnB can prune the true global optimum. Because this bound is the only support for the deterministic global-optimality claims in the abstract and Section 5, it is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a registration cost for affine subspaces (lines and planes) by embedding them into a higher-dimensional Grassmannian and measuring projection residuals of the embedded bases. It derives an SE(n) action on the affine Grassmannian (Theorem 1), proves that zero Grassmann distance is equivalent to zero projection residual (Theorem 2), and formulates a cost function explicitly parameterized by rotation R and translation t (Problem 2, Eq. (7)). The authors then develop a branch-and-bound solver that maximizes inlier sets for rotation and translation and evaluate the method on object registration, RGB-D odometry, perspective-n-line, and correspondence-free localization tasks.","tokens_in":24992,"tokens_out":4729,"duration_ms":49471,"significance":"If the cost function and the bounds were fully validated, the paper would provide a unified geometric framework for line/plane registration that avoids the sign ambiguity of vector parameterizations. The authors give self-contained derivations of the group action and of the zero-set equivalence, release code, and compare against standard baselines on public datasets. The two-stage BnB inlier-set maximization pipeline is a practical contribution. However, the advertised equivalence with geodesic-distance minimization and the deterministic global-optimality guarantee rest on points that need correction or additional proof, so the paper requires major revision.","major_comments":[{"comment":"The abstract and Section 4 claim that Problem 2 (Eq. (7)) minimizes the squared Grassmann geodesic distance of Eq. (6). Appendix 7.4 proves only that the residual in Eq. (31) is zero if and only if the Grassmann distance is zero; it does not establish equality between the summed squared projection residuals in Eq. (7) and the summed squared geodesic distances in Eq. (6). For 1D subspaces, the projection residual is |sin θ| while the geodesic distance is θ, so the two objectives differ away from zero. The phrase 'equivalently optimize Eq. (6)' is therefore unsupported, and the claim of 'minimizing the geodesic distance' should be replaced by a statement that Eq. (7) is a chordal/projection cost with the same zero set, unless a new proof of cost equivalence is supplied.","section":"§4.2, Eq. (7); Appendix 7.4"},{"comment":"The translation lower bound relies on the assertion that Eq. (60) is maximized at a vertex of the translation cube because the inner products (tilde_b2)^T tilde_b1'(R*, t) are 'monotonic increasing, monotonic decreasing, convex, or concave', justified only by 'in every tested case in our experiments'. This is not a proof, and the geometric statement preceding it is inaccurate: t -> R*b1 + R*(I - dd^T)R*^T t maps the cube through a rank-2 linear map, so after normalization the image is a 2D surface in S^3, not a line segment with two endpoints. A concave inner-product profile can place the maximum of Eq. (60) in the interior of the cube, in which case Eq. (62) overestimates the lower bound and the BnB in Algorithm 5 can prune the global optimum. Because this bound is the only support for the deterministic global-optimality claims in the abstract and Section 5, the authors must either prove a vertex-maximum property or replace the bound with a valid one.","section":"Appendix 8.4.1, Eqs. (60)–(63)"}],"minor_comments":[{"comment":"The text references 'Theorem 3' without defining it; presumably Theorem 1 or Corollary 1.1 is intended.","section":"Appendix 8.4.1"},{"comment":"The condition 'If.size < ¯νr' compares an integer inlier count with a real-valued bound; please clarify the intended comparison and notation.","section":"Algorithms 3–5"},{"comment":"Because the method solves Eq. (8) for R and then Eq. (9) for t, the manuscript should state explicitly that the two-stage procedure does not guarantee joint global optimality of Problem 2 unless the cost separates exactly.","section":"§4.2, Eq. (8)–(9)"},{"comment":"The empirical observations that the projected curve length equals the geodesic length and that l1 + l2 = π are reported without proof; either add a derivation or soften the claim.","section":"Appendix 9.2, Fig. 9–10"},{"comment":"The localization experiment reports 0.65° and 0.03% error but gives no number of test frames and no comparison baseline; please add these details.","section":"§5.4, Correspondence-free Localization"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision venue, and the code release is commendable. However, the advertised geodesic-distance minimization and global-optimality guarantees need to be reconciled with the actual cost and bounds; I recommend revision rather than rejection because the projection-residual cost is a plausible registration objective even if it is not the geodesic distance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the explicit cost function in Eq. (7) that expresses the Grassmannian alignment residual as a function of R and t, and the proof that its zero set matches zero geodesic distance. That derivation is correct and useful; it gives a unified objective for line and plane registration without sign ambiguity. The experiments are extensive, the code is available, and the rotation BnB uses standard Go-ICP bounds.\n\nThe soft spots are two. First, the paper says it minimizes the geodesic distance, but the cost is actually the squared chordal distance (sum of squared projection residuals). These agree at zero but are not the same metric; for near-alignment they are close, but the abstract overstates the connection. That is easy to fix by rewriting the claims.\n\nSecond, and more serious: the translation BnB bound in Appendix 8.4.1 is load-bearing and unproven. The paper claims that the set of b1'(R*,t) over the translation cube forms a line segment with endpoints at the vertices. It does not. The map t -> R*(I - dd^T)R*^T t has rank 2, so the image of a 3D cube is a two-dimensional parallelogram. After normalization, it's a 2D surface on the 3-sphere, not an arc. The argument that the inner products are monotonic/convex/concave 'in every tested case' is not a proof, and a norm maximum can sit in the interior of a 2D region. If that happens, Eq. (62) overestimates the lower bound and the BnB can prune the true global optimum. Since the abstract and Section 5 claim deterministic global optimality, this is a real gap.\n\nThe rest of the paper holds up. The zero-equivalence theorem is correct, the rotation bounds are standard, and the empirical results are consistent. The right fix is to either prove the translation bound (maybe with Lipschitz constants on the sphere) or label the translation solver as heuristic and drop the global optimality claim. With that adjustment, the paper is a solid contribution to the registration subfield.\n\nI would bring it to a reading group and would cite it for the cost function. It deserves a serious referee, but the authors should be pushed to be honest about the translation bound.","headline":"A useful new cost function for affine subspace registration, but the translation BnB's global optimality claim rests on an unproven, likely invalid bound.","tokens_in":25513,"tokens_out":3575,"would_cite":true,"duration_ms":34383,"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":"The paper claims to be the first to derive an optimizable cost function for aligning affine subspaces—lines and planes—by minimizing geodesic distance on the Grassmann manifold directly as a function of rotation $\\mathbf{R}$ and…","keywords":["affine Grassmannian","Grassmann manifold","subspace registration","geodesic distance","branch and bound","inlier-set maximization","line-to-plane registration","RGB-D odometry"],"falsifier":"Generate random line correspondences and, for a fixed rotation and a small translation cube, compute the true maximum over $\\mathbf{t}$ of $\\| (\\tilde{b}_2^\\top \\tilde{b}_1'(R^*, t_0)) \\tilde{b}_1'(R^*, t_0) - (\\tilde{b}_2^\\top \\tilde{b}_1'(R^*, t)) \\tilde{b}_1'(R^*, t) \\|$ using dense sampling or local optimization, and compare it with the maximum over the cube's vertices; finding any interior point with a larger value than all vertices would invalidate Eq. (62), making the lower bound unsound and allowing the branch-and-bound search to prune the global optimum.","tokens_in":24527,"feed_emoji":"📐","tokens_out":7514,"duration_ms":69069,"temperature":0.7,"pith_summary":"Existing line and plane registration methods either approximate features with points or parameterize them in Euclidean space, where sign ambiguity (a plane normal versus its negative) distorts the metric. This paper instead works on the affine Grassmannian, the manifold of affine subspaces, and claims the first cost function that expresses the geodesic distance between two Grassmannian features explicitly in terms of a rigid transformation $(\\mathbf{R}, \\mathbf{t})$. The key move is a proof that an embedded subspace's orthonormal basis gives an exact representation of the cost: a smaller subspace lies at zero geodesic distance from a larger one exactly when each of its basis vectors is spanned by the larger subspace's basis. The resulting cost is optimizable and representation-agnostic, and a branch-and-bound extension finds globally optimal inlier sets for 3D line-to-line, line-to-plane, and plane-to-plane registration. The authors demonstrate improved accuracy on object registration, RGB-D odometry, perspective-n-line pose estimation, and correspondence-free localization.","feed_headline":"A Grassmann cost makes line and plane registration globally optimizable","feed_subtitle":"The first optimizable geodesic-distance objective over R and t, backed by a BnB solver that survives heavy outliers.","key_machinery":"The central object is the affine Grassmannian embedding, the map from an affine subspace to a linear subspace one dimension higher, together with the basis-spans-subspace criterion from Theorem 2. The embedding lets a line or plane be represented by an orthonormal matrix whose columns are the transformed source basis and a normalized displacement vector; the Grassmann distance between two such matrices is the metric the paper wants to optimize. The basis criterion replaces the non-differentiable singular-value computation of principal angles with squared Euclidean residuals between each basis vector and its projection onto the transformed target subspace. These residuals are polynomial in $\\mathbf{R}$ and $\\mathbf{t}$, so the cost can be minimized by ordinary nonlinear least squares and bounded for branch-and-bound search.","core_discovery":"Using the orthonormal matrix representation of the affine Grassmannian—embedding a $k$-dimensional affine subspace $A + b_0$ into $\\mathrm{Gr}(k+1, n+1)$ via $Y_{z(A+b_0)} = \\begin{pmatrix} A & b_0/\\sqrt{1+\\|b_0\\|^2} \\\\ 0 & 1/\\sqrt{1+\\|b_0\\|^2} \\end{pmatrix}$—the paper proves that the $\\mathrm{SE}(n)$ group action moves the subspace's linear part and displacement in a closed form (Theorem 1), and that zero Grassmann distance between two embedded subspaces of possibly different dimension is equivalent to every orthonormal basis vector of the smaller one being in the span of the larger one (Theorem 2). From this it constructs Problem 2, Eq. (7): for $N$ paired affine primitives, minimize $\\sum_i \\left( \\sum_{j=1}^{k_i} \\| P_{R \\cdot B_i} a_i^j - a_i^j \\|_2^2 + \\| P_{z(T \\cdot (B_i + d_i^0))} \\tilde{c}_i^0 - \\tilde{c}_i^0 \\|_2^2 \\right)$, where $P$ denotes orthogonal projection onto the transformed source subspace. The paper claims that the zero set of this cost coincides exactly with zero geodesic distance, that the cost separates into rotation-only and translation-dependent terms, and that exact bounds can be derived for a branch-and-bound inlier-set maximizer. Experiments across four computer vision tasks support the claim that this cost improves convergence of existing linear solvers or outperforms them.","pith_inferences":["The basis-spanning criterion is metric-free, so the cost could be generalized to weighted or robustified residuals, folding in feature covariance or correspondence confidence without changing the zero-distance equivalence; the paper does not develop this.","The appendix's numerical observation that the two antipodal straight-line parameter paths map to arcs whose lengths sum to $\\pi$, with one matching the geodesic to numerical precision, hints at an explicit closed-geodesic formula on $\\mathrm{Gr}(2,3)$; if pursued, it could yield a direct differentiable expression for the longer arc connecting two subspaces.","The cost is defined for arbitrary dimensions $n$ and $k$, so it should extend beyond 3D lines and planes to higher-dimensional affine subspace features in structure-from-motion or model fitting; this is a direct generalization of the presented framework."],"forward_implications":["Rotation and translation can be estimated in stages: the linear-subspace terms in Eq. (7) depend only on $\\mathbf{R}$, so an inlier-maximizing rotation search can be run first and the translation search second.","The same cost applies to any pair of affine subspaces of different dimensions, including lines-to-planes, which lets the perspective-n-line problem be reformulated as a line-to-plane registration.","With the derived bounds, the branch-and-bound solver finds a deterministic global optimum of the inlier set, effective at outlier ratios up to 80% in synthetic PnL experiments, and supports correspondence-free localization.","Because the cost is agnostic to the sign of basis vectors, refining a conventional parameter-based linear solver with this cost reduces translation error, as shown in object registration and RGB-D odometry."],"supporting_citations":[{"why":"Supplies the affine Grassmannian formulation for 3D lines and planes that the paper extends; prior work used geodesic distance only for a consistency graph, not as an optimizable pose cost.","marker":"[24]"},{"why":"Defines the affine Grassmannian and its embedding into a Grassmannian of one higher dimension, which is the representation underlying Definition 3.","marker":"[20]"},{"why":"Provides the SVD-based computation of principal angles that the paper replaces with the projection-residual cost.","marker":"[5]"},{"why":"A plane-parameter maximum-likelihood registration solver used as a baseline; its parameter-based metric is the target of the paper's representation-agnostic improvement.","marker":"[11]"},{"why":"A line-registration solver based on Plücker coordinates used as baseline and as initial solution for the proposed refinement.","marker":"[21]"},{"why":"The inlier-set maximization formulation and branch-and-bound structure that the paper adapts for its rotation and translation search.","marker":"[8]"},{"why":"The branch-and-bound Euclidean registration framework and the Space Station data used in the object registration experiment.","marker":"[30]"},{"why":"Supplies the SO(3) cube bound $\\min(\\pi/2, \\sqrt{3}\\,\\sigma_r)$ used in the rotation bounds.","marker":"[40]"}],"fun_headline_variants":["First optimizable geodesic cost for affine subspace registration","Grassmann geodesic distance enables globally optimal subspace alignment","Beyond points: globally optimal registration via Grassmann cost","Branch-and-bound on Grassmann cost beats heavy outlier noise","Optimize rigid transforms on Grassmann manifold for global optimum"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The translation branch-and-bound lower bound rests on the claim that, along the arc traced by the normalized displacement as $\\mathbf{t}$ varies within a cube, the inner product with the fixed embedded target vector is always monotonic, convex, or concave, so the residual difference of Eq. (60) is maximized at a cube vertex; the paper supports this only empirically, writing \"in every tested case in our experiments,\" and gives no proof.","fun_headline_variants_meta":{"raw":{"variants":["First optimizable geodesic cost for affine subspace registration","Grassmann geodesic distance enables globally optimal subspace alignment","Beyond points: globally optimal registration via Grassmann cost","Branch-and-bound on Grassmann cost beats heavy outlier noise","Optimize rigid transforms on Grassmann manifold for global optimum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000902,"raw_usage":{"total_tokens":3981,"prompt_tokens":1145,"completion_tokens":2836,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":761,"completion_tokens_details":{"reasoning_tokens":2757}},"tokens_in":761,"tokens_out":2836,"duration_ms":21532,"temperature":1.0,"reasoning_tokens":2757,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:39:12.373593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate random line correspondences and, for a fixed rotation and a small translation cube, compute the true maximum over $\\mathbf{t}$ of $\\| (\\tilde{b}_2^\\top \\tilde{b}_1'(R^*, t_0)) \\tilde{b}_1'(R^*, t_0) - (\\tilde{b}_2^\\top \\tilde{b}_1'(R^*, t)) \\tilde{b}_1'(R^*, t) \\|$ using dense sampling or local optimization, and compare it with the maximum over the cube's vertices; finding any interior point with a larger value than all vertices would invalidate Eq. (62), making the lower bound unsound and allowing the branch-and-bound search to prune the global optimum.","supporting_citations":[{"cited_title":"Global data association for SLAM with 3D Grassmannian manifold objects","cited_arxiv_id":null,"evidence_quote":"Supplies the affine Grassmannian formulation for 3D lines and planes that the paper extends; prior work used geodesic distance only for a consistency graph, not as an optimizable pose cost."},{"cited_title":"The grass- mannian of affine subspaces","cited_arxiv_id":null,"evidence_quote":"Defines the affine Grassmannian and its embedding into a Grassmannian of one higher dimension, which is the representation underlying Definition 3."},{"cited_title":"Numerical methods for computing angles between linear subspaces","cited_arxiv_id":null,"evidence_quote":"Provides the SVD-based computation of principal angles that the paper replaces with the projection-residual cost."},{"cited_title":"Efficient and accurate registration of point clouds with plane to plane cor- respondences","cited_arxiv_id":null,"evidence_quote":"A plane-parameter maximum-likelihood registration solver used as a baseline; its parameter-based metric is the target of the paper's representation-agnostic improvement."},{"cited_title":"Pluck- ernet: Learn to register 3D line reconstructions","cited_arxiv_id":null,"evidence_quote":"A line-registration solver based on Plücker coordinates used as baseline and as initial solution for the proposed refinement."},{"cited_title":"Globally-optimal inlier set maximisation for camera pose and correspondence estimation","cited_arxiv_id":null,"evidence_quote":"The inlier-set maximization formulation and branch-and-bound structure that the paper adapts for its rotation and translation search."},{"cited_title":"Branch- and-bound methods for Euclidean registration problems","cited_arxiv_id":null,"evidence_quote":"The branch-and-bound Euclidean registration framework and the Space Station data used in the object registration experiment."},{"cited_title":"Go-ICP: A globally optimal solution to 3D ICP point- set registration","cited_arxiv_id":null,"evidence_quote":"Supplies the SO(3) cube bound $\\min(\\pi/2, \\sqrt{3}\\,\\sigma_r)$ used in the rotation bounds."}],"review_version":1}