{"id":"8307463b-9ce9-4203-b52b-8bc5319f6d05","arxiv_id":"1908.09186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Basis point sets encode a point cloud into a fixed-length distance feature, letting a small MLP match PointNet on ModelNet40 with far fewer FLOPs and enabling real-time mesh registration.","lead":"This paper encodes 3D point clouds by measuring the distance from each point in a fixed reference set to the nearest cloud point, producing a fixed-length vector. Simple networks on these vectors match PointNet's classification accuracy with far lower compute, and align body meshes to noisy scans in a single millisecond pass.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first single-pass high-resolution mesh registration' claim is contradicted by the cited 3D-CODED method, which appears to perform exactly this at test time; the novelty claim needs verification.","rationale":"The reader's weakest_assumption focuses on normalization robustness to outliers. That is a legitimate limitation, and the supplementary explicitly admits sensitivity to large outlier clusters. However, it is a robustness caveat, not a falsification of the paper's demonstrated results: the FAUST registration experiment uses the standard FAUST test set, and Dynamic FAUST is presented as a generalization video with acknowledged failure modes. If the concern were the only issue, CONDITIONAL acceptance with a request to discuss robustness would suffice. The 'first single-pass' claim is more load-bearing because it is a binary factual assertion about priority that the paper itself juxtaposes with a citation to 3D-CODED. 3D-CODED's architecture and reported inference protocol (single forward pass, no per-scan optimization) directly match what the paper claims to be first at. If verified, the abstract's second sentence is false; the paper's contribution reduces to applying BPS to a known task, with competitive but not state-of-the-art accuracy (2.327/4.529mm vs 3D-CODED's 1.985/2.878mm). The FLOPs comparison is also worth qualifying (encoding cost is excluded), but that is a quantitative overstatement depending on accounting conventions, not a clean falsification. Thus the priority claim is the most concrete, load-bearing issue. The verdict remains CONDITIONAL: the paper should either substantiate a meaningful distinction between BPS-based registration and 3D-CODED or retract 'first'; the classification contribution is solid enough to survive revision.","tokens_in":14255,"tokens_out":9731,"duration_ms":93766,"concrete_test":"Inspect the test-time inference protocol of 3D-CODED (reference [15] in the paper): run a FAUST scan through the released model and check whether any iterative optimization or per-scan fitting is performed. If the output mesh is produced by a single encoder-decoder forward pass without refinement, then the 'first single-pass high-resolution mesh registration' claim in the abstract and Section 5.2 is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two parts: a classification efficiency result and a registration novelty claim. The second part, 'we present the first method for single-pass high-resolution mesh registration' (Abstract and Sec. 5.2), is directly testable against the paper's own reference [15], 3D-CODED (Groueix et al., ECCV 2018). 3D-CODED is described in the same section as a machine-learning registration method, and its published architecture takes a raw point cloud, computes a global descriptor, and decodes a deformed template mesh in a single feed-forward pass with no per-scan optimization. If that description is accurate, then 3D-CODED is already a single-pass high-resolution mesh registration method, so the 'first' claim is false. The paper's dismissal of [15] as 'multistage models with complex internal representations' does not establish iterative test-time behavior; 'multistage' may refer to training or internal architecture rather than runtime optimization. Because this is an explicit priority/novelty assertion in the abstract, getting it wrong changes the nature of the contribution from 'first' to 'an alternative,' though the BPS representation and its classification efficiency remain intact. The normalization robustness issue raised in the supplementary (large outlier clusters 'drastically change the representation') is a real limitation for cluttered scans, but it is acknowledged by the authors and does not invalidate the FAUST registration experiment; it is secondary to the falsifiable novelty claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes basis point sets (BPS), a fixed-length representation of point clouds obtained by computing the distances from each point in a fixed, arbitrary basis set to the nearest point in the input cloud. The representation is invariant to point ordering by construction and can be fed to standard architectures such as MLPs, 3D convolutional networks, and tree ensembles. The paper evaluates BPS on ModelNet40 classification and FAUST mesh registration. For classification, a simple fully connected network on BPS features reaches 89.0% accuracy with 7.6e5 network FLOPs, compared with PointNet's 89.3% and 4.4e8 FLOPs. For registration, a dense network predicts SMPL vertex positions from the BPS input, achieving 2.327 mm intra-subject and 4.529 mm inter-subject error on FAUST, with inference under 1 ms. The paper claims this is the first single-pass high-resolution mesh registration method and emphasizes the efficiency and generality of the representation.","tokens_in":14522,"tokens_out":3713,"duration_ms":38921,"significance":"If the stated claims hold, BPS is a conceptually simple and broadly applicable representation: it is fixed-length, permutation-invariant, parameter-free at encoding time (aside from the hyperparameters k, r, and basis arrangement), and it works with off-the-shelf machine learning algorithms. The classification experiments are carefully controlled, with ablations over occupancy grids, truncated distance fields, and several basis selection strategies, and the reported accuracy is competitive with PointNet at substantially lower network cost. The registration system is fast and demonstrates a generative use of the encoding. However, the novelty claim of being the first single-pass registration method appears to conflict with the cited 3D-CODED method, and the FAUST accuracy is worse than several cited baselines. The classification efficiency claim also needs clarification, since the reported FLOPs exclude the cost of computing the BPS encoding itself.","major_comments":[{"comment":"The claim of presenting 'the first method for single-pass high-resolution mesh registration' is not supported by the paper's own references. The cited 3D-CODED method of Groueix et al. [15] is a learned method that takes a raw point cloud, computes a global descriptor, and decodes a deformed template mesh in a single feed-forward pass without per-scan optimization, which is exactly a single-pass high-resolution mesh registration method. The statement in Sec. 5.2 that current solutions like [15] rely on 'multistage models with complex internal representations' does not establish that they perform iterative test-time optimization; 'multistage' may refer to training or internal architecture rather than runtime behavior. This priority claim should be removed or carefully qualified, and the contribution repositioned as an alternative single-pass approach with a particularly simple representation.","section":"Abstract and Sec. 5.2"},{"comment":"The FAUST results are not just 'comparable but higher' than the state of the art: in the intra-subject challenge the method is worse than 3D-CODED (2.327 mm vs. 1.985 mm) and Stitched Puppets (1.568 mm), and in the inter-subject challenge it is worse than several methods including 3D-CODED, Stitched Puppets, FARM, and Deep Functional Maps. The speed advantage is real and should be the primary claim, but the accuracy comparison should be reported without the implication of near-state-of-the-art accuracy. Please state explicitly that the method trades accuracy for speed.","section":"Sec. 5.2, Table 3"},{"comment":"The headline 'three orders of magnitude less floating-point operations' refers only to the network forward pass, not to the full classification pipeline. The BPS encoding itself requires O(n log n + k log n) operations per cloud (or O(kn) in the direct GPU implementation), and the supplementary material reports encoding times that are real-time but not zero. Since Table 2 lists only network FLOPs, the efficiency comparison is incomplete for a user who must encode every test cloud before classification. Please report the encoding cost separately and state clearly whether the claimed orders-of-magnitude advantage holds end-to-end; if it does not, the abstract and conclusion should be reworded.","section":"Sec. 5.1, Table 2, and Supplementary Sec. 1"}],"minor_comments":[{"comment":"The normalization in Eq. (2) re-centers each cloud by its centroid and rescales by the maximum distance; the supplementary material admits that large clusters of outlier points far from the body 'drastically change the representation due to the size normalization.' This is an acknowledged limitation, but it should be discussed in the main text as a scope restriction for the registration system, since real scans may contain such outliers.","section":"Eq. (2) and Supplementary Sec. 4"},{"comment":"The difference between BPS-MLP (89.0%) and PointNet (89.3%) is within the reported variance of similar models, so the phrase 'reaching the performance of PointNet' is appropriate, but the paper should avoid appearing to claim superiority on this comparison.","section":"Sec. 5.1, Table 2"},{"comment":"There is a typo in the training details: 'dropout (with probability 0.8 and 0.4 repsectively)' should be 'respectively'.","section":"Supplementary Sec. 3"},{"comment":"The caption says 'ModelNe40' but should read 'ModelNet40'.","section":"Fig. 4 caption"},{"comment":"The notation in Eq. (5) uses a set-builder expression with braces but the result is a matrix in R^{k x d}; please make the notation unambiguous, for example by writing the rows explicitly or using a vector of differences.","section":"Sec. 3, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The central BPS representation and the classification experiments are sound and useful. The main problem is the overstated registration novelty claim, which is directly contradicted by the cited 3D-CODED work; this needs to be corrected before publication. I recommend requiring the authors to reposition the registration contribution and to clarify the FLOPs comparison, rather than rejecting the paper, since the core representation and its evaluation are presented cleanly and the issues are fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading for the BPS encoding, but the abstract overstates priority. The 'first single-pass high-resolution mesh registration' claim is not supported: their own reference [15], 3D-CODED, is a single feed-forward deep deformation method that maps a raw scan to a deformed template mesh without test-time optimization. Unless there is some architectural difference that makes 3D-CODED not single-pass in the relevant sense, that claim should be dropped or qualified. The stress-test concern lands.\n\nWhat is genuinely useful: a fixed-length representation of point clouds as distances to a fixed random basis set, compatible with off-the-shelf MLPs, trees, and other standard machinery. The encoding is simple and fast, and the paper evaluates it carefully—comparisons against occupancy grids and truncated distance fields, four basis-selection strategies, reconstruction fidelity, and two downstream tasks. They ship code. The classification result (89.0% vs PointNet's 89.3% on ModelNet40, with far fewer FLOPs for the network) is interesting, though the accuracy difference is within noise and the FLOP count excludes encoding cost. The registration system is genuinely fast—sub-millisecond CPU inference—and its FAUST errors are comparable to much slower state-of-the-art methods, albeit slightly worse than both 3D-CODED and Stitched Puppets.\n\nSoft spots, in order: the priority claim is the biggest. Second, the 'three orders of magnitude' headline applies to the forward pass only; BPS encoding itself costs O(k log n) per cloud, and while their supplementary shows real-time encoding, the full-pipeline efficiency picture is not as stark as the abstract implies. Third, the normalization in Eq. 2 is fragile: the supplementary admits large outlier clusters 'drastically change the representation,' which limits robustness on real scans. The authors acknowledge this, and it does not invalidate the FAUST experiment, but it is a real limitation for the practical use case they advertise.\n\nWho is this for: researchers and engineers working on efficient point cloud learning, CPU-only or embedded 3D understanding, or anyone considering low-cost representations for downstream tasks. The paper deserves serious peer review. A revise-and-resubmit would be appropriate if the priority claim is corrected and the efficiency accounting is made honest.","headline":"Useful practical encoding; the 'first single-pass registration' claim overreaches against the paper's own cited prior work, and the FLOPs accounting needs qualification, but the core contribution is sound.","tokens_in":15059,"tokens_out":1987,"would_cite":true,"duration_ms":20747,"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 point cloud can be encoded as distances from a fixed set of basis points, letting simple networks match specialized architectures at a fraction of the compute.","keywords":["basis point sets","point cloud encoding","shape classification","mesh registration","fixed-length representation","nearest neighbor distances","computational efficiency","fully connected networks"],"falsifier":"Take a trained BPS registration network and one clean test scan; compute the BPS vector once normally and once after appending a large cluster of far-away outlier points that shrink the body under Eq. 2 normalization. If the predicted mesh changes dramatically or the ModelNet40 classifier's accuracy drops, the robustness assumption collapses; conversely, if the output is stable, the main practical limitation the supplementary material concedes is resolved.","tokens_in":1830,"feed_emoji":"🧊","tokens_out":3218,"duration_ms":88505,"temperature":0.7,"pith_summary":"The paper introduces basis point sets (BPS), a way to turn a point cloud of arbitrary size and ordering into a fixed-length vector: for each of $k$ fixed basis points, record the distance to the nearest point in the cloud. Because the resulting vector has a fixed length and a fixed order, it can be fed directly into ordinary fully connected networks, boosted trees, or other standard machine learning methods without specialized point-cloud layers. On the ModelNet40 shape classification benchmark, a simple MLP trained on BPS features matches PointNet's accuracy while using roughly three orders of magnitude fewer floating-point operations. For mesh registration, the same encoding lets a network predict a high-resolution registered mesh from a noisy 3D scan in a single forward pass, under one millisecond on a CPU, with no per-scan optimization. The paper argues that BPS is a general, efficient alternative to occupancy grids and direct point-cloud networks.","feed_headline":"Fixed-point distances make 3D scans easy for ordinary neural networks","feed_subtitle":"Matching PointNet's accuracy at 0.2% of the compute, and real-time mesh fitting from noisy scans.","key_machinery":"The central object is the basis point set encoding: a fixed, arbitrary set of $k$ points sampled inside a ball, where each basis point $b$ contributes the minimal Euclidean distance from $b$ to any point in the normalized cloud (Eq. 4), optionally extended to the full delta vector to the nearest point (Eq. 5). The fixed basis gives the representation a fixed length and a fixed ordering even though the cloud is unordered, and the use of global, unsigned, non-truncated distances means every entry carries information about the whole shape while the vector varies smoothly under transformations. This carries the argument because it converts the core difficulties of point clouds - variable cardinality and lack of order - into a standard fixed-dimension input, so the efficiency gains come from replacing specialized networks with simple dense layers.","core_discovery":"The central claim is that a point cloud's geometry can be captured for learning purposes by a residual representation: minimal distances from each member of a fixed basis point set to the cloud, optionally with the full delta vector to the nearest point. This representation accepts an unordered point cloud of any size and returns an ordered, fixed-length feature vector, and it preserves surface details far better than occupancy grids of equivalent encoding length. Using these features, a plain fully connected network reaches PointNet-level classification accuracy on ModelNet40 with about $7.6\\times10^5$ floating-point operations compared with PointNet's $4.4\\times10^8$, and a similarly simple regressor maps noisy scans to FAUST-registered meshes in one pass, with errors of 2.327 mm intra-subject and 4.529 mm inter-subject and inference under 1 ms. The paper's position is that BPS is a preprocessing step that lets standard machine learning algorithms handle point clouds directly, sidestepping the need for custom permutation-invariant architectures.","pith_inferences":["Because BPS is a generic fixed-length representation, it could serve as a shared input layer for multi-task models that jointly classify, segment, and register scans, a combination the paper does not explore.","The revealed sensitivity to normalization suggests a testable extension: augmenting training with outlier-heavy point clouds or replacing maximum-distance rescaling with a robust scale estimator could make the method usable on raw scanner output with floor and background clutter.","The success of randomly placed basis points indicates that the specific geometry of the basis matters less than the distance-to-cloud statistic itself, so learned or adaptively placed basis points might further improve fidelity without changing the architecture.","The same encoding could be plugged into generative architectures such as denoising diffusion models for point cloud generation, an application beyond the discriminative and regression tasks demonstrated here."],"forward_implications":["A point cloud can be processed by any machine learning algorithm that accepts fixed-length vectors, including fully connected networks, random forests, and gradient-boosted trees, without custom permutation-invariant layers.","On ModelNet40, BPS features with a simple MLP reach 89.0% accuracy, matching PointNet, while using about 0.2% of PointNet's floating-point operations; a BPS grid with 3D convolutions reaches 90.8%, matching PointNet++.","Mesh registration to noisy scans can be done in a single feed-forward pass in under 1 ms on a CPU, enabling real-time alignment without iterating per scan.","At the same encoding length, BPS reconstructs surface details with less than half the Chamfer error of occupancy grids, and its reconstruction quality is close to that of subsampled raw point clouds while remaining fixed-length and ordered.","Encoding complexity is $O(n\\log n + k\\log n)$ per cloud with ball-tree nearest-neighbor search, so real-time encoding is practical for clouds with $n < 10^5$ points and $k < 10^4$ basis points."],"supporting_citations":[{"why":"Supplies the PointNet baseline that BPS-MLP matches in classification accuracy while using fewer FLOPs.","marker":"[34]"},{"why":"Provides the ModelNet40 dataset and the point-sampling protocol used for classification and reconstruction experiments.","marker":"[46]"},{"why":"Provides the VoxNet occupancy-grid baseline and the 3D ConvNet architecture adapted for BPS grids.","marker":"[29]"},{"why":"Supplies the FAUST benchmark used to evaluate single-pass mesh registration.","marker":"[2]"},{"why":"Provides the SMPL body model used to generate posed, shaped training scans for the registration network.","marker":"[27]"},{"why":"Is the main registration baseline whose multistage optimization BPS replaces with a single pass.","marker":"[15]"},{"why":"Defines the truncated signed distance field representation that BPS generalizes and compares against.","marker":"[8]"},{"why":"Supplies the ball-tree data structure behind the claimed O(n log n + k log n) encoding complexity.","marker":"[32]"},{"why":"Is the iterative stitched-puppets registration baseline that BPS outperforms in speed by orders of magnitude.","marker":"[49]"}],"fun_headline_variants":["Basis point sets: standard NNs handle point clouds with 0.2% compute","Distance-to-fixed-points representation cuts point cloud learning compute 1000x","Basis point sets enable real-time mesh registration from noisy scans","PointNet-level accuracy from plain fully connected nets via basis point sets","BPS: ordered features for unordered clouds, 0.2% of PointNet's FLOPs"],"cache_read_input_tokens":17152,"weakest_assumption_plain":"The load-bearing premise is that the Eq. 2 normalization - centering each cloud at its centroid and rescaling by the maximum distance - is stable, so a scan with a large cluster of distant outlier points does not produce a BPS vector far outside the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Basis point sets: standard NNs handle point clouds with 0.2% compute","Distance-to-fixed-points representation cuts point cloud learning compute 1000x","Basis point sets enable real-time mesh registration from noisy scans","PointNet-level accuracy from plain fully connected nets via basis point sets","BPS: ordered features for unordered clouds, 0.2% of PointNet's FLOPs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3230,"prompt_tokens":984,"completion_tokens":2246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2140}},"tokens_in":600,"tokens_out":2246,"duration_ms":15015,"temperature":1.0,"reasoning_tokens":2140,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:19:12.750997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained BPS registration network and one clean test scan; compute the BPS vector once normally and once after appending a large cluster of far-away outlier points that shrink the body under Eq. 2 normalization. If the predicted mesh changes dramatically or the ModelNet40 classifier's accuracy drops, the robustness assumption collapses; conversely, if the output is stable, the main practical limitation the supplementary material concedes is resolved.","supporting_citations":[{"cited_title":"Pointnet: Deep learning on point sets for 3d classification and segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the PointNet baseline that BPS-MLP matches in classification accuracy while using fewer FLOPs."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"Provides the ModelNet40 dataset and the point-sampling protocol used for classification and reconstruction experiments."},{"cited_title":"Voxnet: A 3d convolutional neural network for real-time object recognition","cited_arxiv_id":null,"evidence_quote":"Provides the VoxNet occupancy-grid baseline and the 3D ConvNet architecture adapted for BPS grids."},{"cited_title":"Faust: Dataset and evaluation for 3d mesh registration","cited_arxiv_id":null,"evidence_quote":"Supplies the FAUST benchmark used to evaluate single-pass mesh registration."},{"cited_title":"Smpl: A skinned multi-person linear model","cited_arxiv_id":null,"evidence_quote":"Provides the SMPL body model used to generate posed, shaped training scans for the registration network."},{"cited_title":"3d-coded : 3d correspondences by deep deformation","cited_arxiv_id":null,"evidence_quote":"Is the main registration baseline whose multistage optimization BPS replaces with a single pass."},{"cited_title":"A volumetric method for building complex models from range images","cited_arxiv_id":null,"evidence_quote":"Defines the truncated signed distance field representation that BPS generalizes and compares against."},{"cited_title":"Five balltree construction algorithms","cited_arxiv_id":null,"evidence_quote":"Supplies the ball-tree data structure behind the claimed O(n log n + k log n) encoding complexity."},{"cited_title":"The stitched puppet: A graphical model of 3d human shape and pose","cited_arxiv_id":null,"evidence_quote":"Is the iterative stitched-puppets registration baseline that BPS outperforms in speed by orders of magnitude."}],"review_version":1}