{"id":"4c96f53e-67fa-4e5b-abe6-54dfe6e917ff","arxiv_id":"2412.00952","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ESCAPE achieves rotation-equivariant 3D shape completion by encoding shapes as distances to rotation-covariant anchor points and decoding predicted distances back into coordinates via optimization.","lead":"This paper introduces ESCAPE, a 3D shape completion method that represents a partial object by distances to a few selected anchor points, then uses a transformer to predict distances for the completed shape and an optimizer to recover point coordinates. Because the distance representation is rotation-invariant and the anchors rotate with the input, the completed shape stays aligned with the input without any pose estimation module.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Coordinate optimization has a mirror ambiguity when anchors are coplanar, so exact rotation equivariance is not guaranteed for planar or near-planar partial inputs.","rationale":"The reader's weakest assumption was FPS tie-breaking among equidistant points. That concern is weaker than it appears: if the input point order is preserved under rotation, an index-based argmax tie-break is exactly rotation-equivariant, because rotation preserves all pairwise distances and hence the tie set. The more serious gap is downstream of anchor selection: the optimization that converts predicted distances into coordinates is only rotation-equivariant if the distance-to-anchor constraints uniquely determine each point. Uniqueness requires the anchors to affinely span R^3, and the paper provides no argument that FPS on partial depth scans, especially for planar objects, produces such anchors. When the anchors are coplanar, Eq. (5) has a mirror-pair of solutions and the centroid initialization is a critical point in the normal direction, so the branch is selected by numerical accident rather than by the input rotation. The paper's own Section F limitation statement points at the optimization procedure as a source of non-rotation-invariance, which is consistent with this concern. I still regard the overall claim as conditionally plausible: the distance-feature idea is sound, and a direct test on near-coplanar anchor sets could either confirm that the ambiguity is benign in practice or reveal a genuine failure mode. Because the existing verdict already conditions acceptance on resolving such robustness issues, I leave it unchanged.","tokens_in":16756,"tokens_out":9200,"duration_ms":109672,"concrete_test":"Construct a synthetic flat partial input (points exactly on a plane) plus a complete target with off-plane points, and run ESCAPE's anchor selection on it and on R(input) for several rotations. Compute the centered anchor covariance; record the smallest singular value. Then solve Eq. (5) with the same predicted distance matrix twice: once initialized at the anchor centroid and once at the centroid plus a small normal perturbation, for both canonical and rotated anchors. If the two initializations yield different completions with near-equal loss, or if the canonical and rotated optimizer outputs differ by more than the expected rigid rotation (e.g., the outputs are related by an additional reflection across the anchor plane), the central equivariance claim fails for this regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claimed equivariance needs three links: anchor selection is rotation-equivariant, the predicted distance matrix is rotation-invariant, and the coordinate optimization is rotation-equivariant. The paper's reconstruction uniqueness argument (Theorem 1, Section 3.3) requires anchors in general position, but the FPS-plus-curvature selection from a single partial depth scan does not guarantee this. For planar or near-planar surfaces (tables, desks, flat panels, many single-view scans), all selected anchors can lie in or very near a single plane. In that regime, distances to anchors determine any completed point only up to reflection across the anchor plane, so Eq. (5) has two equally valid solutions for each point. The optimization is initialized at the centroid of the anchors, which lies in that plane; the normal-direction Jacobian of the distance residuals is zero there, so Levenberg-Marquardt can stall at the plane or break the symmetry by an arbitrary numerical decision. Because that branch choice is not coupled to the input rotation, the final coordinates are not guaranteed to satisfy F(RP) = R F(P). This is independent of FPS tie-breaking: even a perfectly equivariant anchor selector leaves the reflection ambiguity unresolved. The paper itself flags the optimization as a limitation in Section F, and Section 3.3's phrase 'unique solution (up to reflection)' conflates the general-position case with the coplanar case. Thus the load-bearing premise that the entire pipeline is rotation-equivariant for arbitrary inputs is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"ESCAPE proposes a rotation-equivariant point cloud completion pipeline. It selects k anchor points from the partial input via deterministic farthest point sampling followed by curvature-based refinement, represents all points by their Euclidean distances to the anchors (Eq. 1), predicts the distance matrix of the completed cloud with a transformer modified from AdaPoinTr, and finally recovers point coordinates by Levenberg-Marquardt optimization (Eq. 5). Experiments on PCN, OmniObject3D, and KITTI report large improvements over PCA-aligned and ConDor-canonicalized non-equivariant baselines on rotated inputs, together with ablations on anchor selection, noise robustness, and partiality. The appendix includes theoretical error-bound arguments, additional baseline comparisons, and a limitations section.","tokens_in":17042,"tokens_out":4981,"duration_ms":46274,"significance":"If the central equivariance claim were established exactly, ESCAPE would be a practically significant contribution: it avoids pose estimation modules and achieves a large empirical gap on rotated PCN inputs (CD-L1 10.58 vs. 26.65 for the best PCA-aligned baseline). The paper also provides useful baselines (SCARP, Vector Neurons, PPF-Snowflake, ConDor), real-world evaluations, and an ablation study on anchor selection. However, the current theoretical guarantees are overclaimed: the constant-error-bound theorem applies only to the distance map, not to the end-to-end completion, and the coordinate optimization has a reflection ambiguity for non-general-position anchors that breaks exact equivariance. The empirical results remain valuable, but the paper needs to either prove a qualified equivariance statement or substantially narrow the claim.","major_comments":[{"comment":"The end-to-end equivariance claim requires that the solution of Eq. (5) be rotation-equivariant. Theorem 1 guarantees uniqueness only when the anchors are in general position, but the anchor selection of Section 3.1 does not enforce this condition. For planar or near-planar partial scans (e.g., tables, desks, and many single-view depth maps), all k anchors can lie in or very near a single plane. In that regime the distances to the anchors determine each predicted point only up to reflection across the anchor plane, and the Levenberg-Marquardt initialization at the centroid of the anchors lies on that plane, where the normal-direction residual Jacobian is zero. The branch choice is therefore not coupled to the input rotation, so F(RP) = R F(P) is not guaranteed. Section F itself states that the optimization 'prevents the model from being rotation-invariant,' which contradicts the unqualified claim in Sections 3.3 and 4.2. Please add an explicit equivariance-error experiment on planar categories under in-plane rotations, and either prove an approximate-equivariance bound or restrict the claim to inputs whose selected anchors are in general position.","section":"Section 3.3, Eq. (5), and Section F"},{"comment":"Theorem A.1 proves a Lipschitz bound on the input distance matrix, |d(p+epsilon,a)-d(p,a)| <= ||epsilon||, but it does not bound the error of the predicted completed point cloud. The end-to-end error additionally depends on the transformer's prediction error and on the optimization in Eq. (5). Calling this a constant error bound O(1) independent of network depth is therefore not supported for the full pipeline. In addition, the comparison with Vector Neurons in Eq. (8) is not parameter-free, since alpha depends on learned weight norms and on the rotation matrices; a fair comparison would require a bound on those norms or matched training conditions. Please state precisely which quantity enjoys the O(1) bound and which components of the pipeline are excluded.","section":"Appendix A.1, Eq. (10)"},{"comment":"Appendix B.1 claims that initializing FPS from the centroid gives 'exactly same results for the same input independent of rotation,' but no tie-breaking rule is specified. For inputs with symmetries, multiple points can be equidistant from the centroid or from the already-selected set, and a list-order tie-break is not rotation-equivariant; after rotation a different anchor set can be selected. The curvature refinement in Section 3.1 has a related ambiguity, because PCA-based curvature (Eqs. 3-4) is computed from eigendecompositions whose eigenvector signs and order are not canonically defined. Please specify a canonical tie-breaking rule, prove invariance for generic point clouds, and report a test on symmetric or near-symmetric shapes, since exact equivariance of the anchor selector is a load-bearing premise for the distance features.","section":"Appendix B.1"},{"comment":"The theorem as stated says the distance matrix uniquely determines P up to rigid transformation, while Section 3.3 concedes the optimization has a unique solution only 'up to reflection.' These are different statements, since a reflection is not a rigid motion in SO(3), and the proof sketch's 'intersection of k spheres in general position yields a unique point' does not rigorously cover the n-point case or the coplanar-anchor case. Because reconstruction uniqueness is load-bearing for the coordinate optimization, please provide a complete proof with the precise general-position conditions and state clearly when only uniqueness up to reflection holds.","section":"Theorem 1 and Section 3.3"}],"minor_comments":[{"comment":"The claim that ESCAPE is 'the only model where prediction is unaffected by the input rotation' is based on a single random rotation; please report results over multiple rotation seeds, e.g., mean and standard deviation, to support the invariance statement empirically.","section":"Section 4.2, Table 1"},{"comment":"The phrase 'prevents the model from being rotation-invariant' should be distinguished from 'rotation-equivariant'; as written it appears to contradict Section 3.3 and should be clarified or reconciled.","section":"Section F"},{"comment":"There is a typo in the contributions list: 'We present the an end-to-end rotation-equivariant shape completion method.'","section":"Section 1"},{"comment":"Reference [53] duplicates reference [52], and Table 3 uses 'MMID' where the metric is presumably MMD; please correct the notation.","section":"References and Table 3"},{"comment":"The ablation in Table 5 reports a CD-L1 of 14.74 for plain FPS on 'a subset of the PCN dataset,' while Table 4 reports an average of 10.58 for the full method; the subset size, composition, and relation to the main result should be specified.","section":"Table 5 and Table 4"},{"comment":"The symbol R is used both for a rotation matrix and for a layer output; please use distinct notation for clarity.","section":"Appendix A.1, Eqs. (7)-(8)"}],"recommendation":"major_revision","confidential_remarks":"The paper is borderline between major revision and reject because the unqualified exact-equivariance claim is contradicted by the paper's own limitation section. I recommend major revision rather than reject because the empirical contribution is substantial and the claim can be repaired by narrowing the scope, adding an equivariance-error evaluation, and providing the missing tie-breaking and general-position analysis. Please also verify that the anchor-encoding idea is positioned sufficiently against prior learned-keypoint and distance-distribution descriptors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read ESCAPE with the stress-test note in hand, and the stress-test holds. For flat or near-planar partial inputs, the FPS-plus-curvature anchor selection can return anchors that all lie in one plane. In that regime the distance constraints determine each completed point only up to reflection across the anchor plane. The optimization is initialized at the anchor centroid, which sits in that plane, so the normal-direction Jacobian of the distance residuals is zero there. Levenberg-Marquardt then has no rotation-coupled reason to choose one side of the plane. That breaks F(RP) = R F(P) exactly, independent of FPS tie-breaking. Section 3.3's phrase \"unique solution (up to reflection)\" is where the gap sits, and the paper's own Section F admits the optimization is a limitation. So the exact equivariance guarantee for arbitrary inputs is not established.\n\nThat said, this is a legitimate method contribution. Distance-to-anchor encoding, an AdaPoinTr-style transformer, curvature-refined FPS anchors, and coordinate optimization together form a new pipeline for rotation-aware completion. The rotated PCN numbers are strong: 10.58 CD-L1 versus 26.65 for the best PCA-aligned baseline, and the ConDor canonicalization baseline in the appendix makes the comparison fairer. The authors also honestly report that canonical-input performance is worse than the baselines (10.58 vs 6.53) and disclose the OmniObject median protocol. No fitted parameters are dressed as predictions; training is supervised on standard splits. The self-citations to Riga and Rotation-Invariant Transformer appear in related work and the central claim does not rest on those papers.\n\nThe soft spots are real but addressable. The theory section overclaims: Theorem 1's proof sketch says intersection of spheres gives a unique point, but the \"up to reflection\" caveat in Section 3.3 is the correct statement only when anchors are coplanar, and the anchor selector can produce exactly that case. The deterministic FPS tie-breaking for equidistant points is not specified either. No code or data splits are released, which hurts reproducibility. These are not fatal to the empirical story, but they do mean the paper's central theoretical claim needs to be scoped down.\n\nThis paper is for people working on point cloud completion in unposed or robotics settings. It deserves a serious referee. I would send it to review, asking for code, a corrected equivariance statement, and either a tie-breaking specification or an empirical study of degenerate anchor configurations.","headline":"Useful empirical entry on rotation-equivariant completion, but the exact equivariance claim outruns the theory: coplanar anchors leave a reflection ambiguity the paper does not resolve.","tokens_in":17544,"tokens_out":3125,"would_cite":true,"duration_ms":32190,"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":"ESCAPE claims that a transformer over anchor-point distances, followed by coordinate optimization, completes partially scanned 3D shapes in arbitrary rotations without pose estimation.","keywords":["rotation-equivariant shape completion","anchor point encoding","distance-based 3D representation","point cloud transformer","arbitrary pose completion","furthest point sampling","Chamfer distance","real-world point cloud completion"],"falsifier":"Take a point cloud with an exact rotational symmetry, such as a uniform sphere or a cube sampled symmetrically, rotate it by an angle that permutes the equidistant candidates for one anchor, and compare the ESCAPE anchor set and completion for the two inputs; if the anchors or the completed clouds differ, the claimed exact rotation equivariance fails. A numerical version is to measure CD-L1 between completions of a symmetric input and its rotated copy across many random rotations and check whether every sample gives zero difference.","tokens_in":16581,"feed_emoji":"🔄","tokens_out":10589,"duration_ms":88073,"temperature":0.7,"pith_summary":"ESCAPE is a point-cloud completion method designed so that the answer is correct no matter how the partial input is rotated, without ever estimating the object's pose. It represents every input point by its distances to a small set of anchor points chosen from the cloud, feeds that rotation-invariant distance matrix to a transformer, predicts a distance matrix for the completed shape, and then solves for point coordinates. On rotated versions of the PCN benchmark the method reports Chamfer distance 10.58, while the strongest PCA-aligned baseline degrades to 26.65; on the OmniObject real-world benchmark it reports 18.82 versus 29.56 for the best comparison. The paper argues that the distance representation keeps reconstruction error constant with network depth and gives unique reconstructions once enough anchors are available. A sympathetic reader would take away that pose-free completion is achievable by replacing coordinates with anchor distances, at the price of slightly worse results on canonical inputs.","feed_headline":"Distances to anchor points finish 3D shapes in any pose","feed_subtitle":"Trained on canonical inputs only, ESCAPE holds Chamfer-L1 at 10.58 under rotation while PCA-aligned baselines fall to 26.65.","key_machinery":"The load-bearing object is the distance matrix $D \\in \\mathbb{R}^{n \\times k}$ between the $n$ input points and $k = 8$ anchor points selected from the partial cloud by a deterministic farthest-point sampling initialized at the centroid, then refined within clusters to points of highest estimated curvature. Because distances are rotation-invariant while the anchor coordinates rotate with the input, a transformer that consumes only distances predicts a complete-shape distance matrix $\\hat{D}_c$ that is also rotation-invariant; the final completed cloud is recovered by minimizing $\\sum_j (\\|p - a_j\\|_2 - \\hat{d}_{ij})^2$ per point with Levenberg-Marquardt, and the recovered coordinates inherit the input rotation. The paper backs this with a reconstruction-uniqueness theorem ($k \\ge 4$ anchors in general position determine each point up to reflection) and an error-bound argument showing perturbations shift distances by at most the perturbation norm, independent of network depth.","core_discovery":"The central claim is that rotation-equivariant shape completion can be achieved without specialized equivariant layers, pose estimation, or rotation augmentation, by encoding shapes as distances to rotation-covariant anchor points. The paper states that with its deterministic anchor selection and distance-based transformer, the final coordinates predicted by the pipeline retain the same orientation as the partial input cloud, making ESCAPE the only model where prediction is unaffected by the input rotation. Concretely, training only on canonical PCN data and testing on randomly rotated inputs yields CD-L1 of 8.14 to 13.86 per category, averaging 10.58, whereas PCA-aligned transformer baselines range from 26.65 to 92.15; the same weights transfer to unknown-pose OmniObject depth scans and rotated KITTI LiDAR cars. The paper also claims two theoretical properties of the distance representation: identical distance encodings imply isometric shapes, and input perturbations change the distance matrix by at most the perturbation size, a constant error bound independent of network depth.","pith_inferences":["Editorial inference: because the network sees only distances, the same ESCAPE weights can in principle complete shapes at varying levels of partiality without retraining, since the anchors adapt to whichever points are visible.","Editorial inference: the anchor-point mechanism is modular, so replacing curvature-based anchors with any rotation-equivariant keypoint detector is a direct extension that the paper's design already permits.","Editorial inference: the constant-error property suggests ESCAPE-style encodings could help other pose-free point-cloud tasks, such as registration or object detection, where deep equivariant networks currently suffer error accumulation."],"forward_implications":["A completion model trained once on canonical data can be deployed on arbitrarily rotated scans without a pose-estimation preprocessing stage.","Because equivariance comes from the encoding rather than from custom layers, the same anchor-distance representation could be dropped into other point-cloud transformer architectures.","The constant error bound implies that deeper networks built on distance encodings should not accumulate rotational or positional error, unlike layerwise equivariant features.","The method accepts a canonical-input trade-off: existing aligned baselines score 6.53 to 8.38 on canonical PCN inputs while ESCAPE scores 10.58, meaning the equivariance guarantee is bought with some accuracy on perfectly aligned data."],"supporting_citations":[{"why":"Provides the PCN dataset and train/test protocol that define the main rotated-input benchmark and the canonical training data.","marker":"[50]"},{"why":"Supplies the AdaPoinTr transformer architecture that ESCAPE modifies to consume distance-to-anchor features instead of coordinates.","marker":"[49]"},{"why":"Vector Neurons is the equivariant framework whose layerwise error growth O(alpha^L) ESCAPE contrasts with its constant O(1) distance-matrix error bound.","marker":"[5]"},{"why":"SCARP is the existing arbitrary-pose completion baseline that ESCAPE must beat and the source of the learned-canonicalization setup.","marker":"[32]"},{"why":"Osada et al.'s shape distributions motivate representing shape geometry by distances between points, the conceptual basis of anchor-distance encoding.","marker":"[22]"},{"why":"ConDor is used to canonicalize rotated inputs for non-equivariant baselines, establishing the comparison that ESCAPE outperforms.","marker":"[29]"}],"fun_headline_variants":["Anchor-point distances make shape completion rotation-proof","No pose estimation: ESCAPE finishes shapes in any rotation","Simple distance encoding gives equivariant 3D completion","Outperforms PCA baselines under rotation by 2.5x","Anchor-point encoding: the only rotation-equivariant completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole equivariance chain rests on the anchor-selection procedure being exactly rotation-equivariant for every input, but the paper's deterministic farthest-point sampling does not specify how ties among equidistant points are broken; under rotation, a symmetric or near-symmetric cloud can therefore produce different anchors, different distance features, and a different completion.","fun_headline_variants_meta":{"raw":{"variants":["Anchor-point distances make shape completion rotation-proof","No pose estimation: ESCAPE finishes shapes in any rotation","Simple distance encoding gives equivariant 3D completion","Outperforms PCA baselines under rotation by 2.5x","Anchor-point encoding: the only rotation-equivariant completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000242,"raw_usage":{"total_tokens":1518,"prompt_tokens":934,"completion_tokens":584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":550,"tokens_out":584,"duration_ms":5964,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:50:28.444189+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a point cloud with an exact rotational symmetry, such as a uniform sphere or a cube sampled symmetrically, rotate it by an angle that permutes the equidistant candidates for one anchor, and compare the ESCAPE anchor set and completion for the two inputs; if the anchors or the completed clouds differ, the claimed exact rotation equivariance fails. A numerical version is to measure CD-L1 between completions of a symmetric input and its rotated copy across many random rotations and check whether every sample gives zero difference.","supporting_citations":[{"cited_title":"In: 2018 international conference on 3D vision (3DV)","cited_arxiv_id":null,"evidence_quote":"Provides the PCN dataset and train/test protocol that define the main rotated-input benchmark and the canonical training data."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence (2023) 3, 4, 6, 7, 14","cited_arxiv_id":null,"evidence_quote":"Supplies the AdaPoinTr transformer architecture that ESCAPE modifies to consume distance-to-anchor features instead of coordinates."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Vector Neurons is the equivariant framework whose layerwise error growth O(alpha^L) ESCAPE contrasts with its constant O(1) distance-matrix error bound."},{"cited_title":"In: 2023 IEEE International Conference on Robotics and Automation (ICRA)","cited_arxiv_id":null,"evidence_quote":"SCARP is the existing arbitrary-pose completion baseline that ESCAPE must beat and the source of the learned-canonicalization setup."},{"cited_title":"ACM Transactions on Graphics (TOG) 21(4), 807–832 (2002) 2","cited_arxiv_id":null,"evidence_quote":"Osada et al.'s shape distributions motivate representing shape geometry by distances between points, the conceptual basis of anchor-distance encoding."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"ConDor is used to canonicalize rotated inputs for non-equivariant baselines, establishing the comparison that ESCAPE outperforms."}],"review_version":1}