{"id":"89ff6050-afda-420d-a343-7c765765d7fa","arxiv_id":"1908.04512","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new point cloud convolution that interpolates features to fixed discrete kernel weights with density normalization achieves state-of-the-art on ModelNet40, ShapeNet Parts, and S3DIS.","lead":"This paper introduces InterpConv, a convolution operation that works directly on unordered 3D point clouds by interpolating point features onto discrete kernel weights. The authors show that networks built from this layer achieve state-of-the-art accuracy on three standard 3D recognition benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sparsity invariance is asserted but never demonstrated; the normalization only makes aggregation a weighted average over a hard-cutoff neighborhood, which remains density-dependent for finite samples.","rationale":"The single most load-bearing concern is the untested sparsity-invariance claim, because it is a stated central contribution and the papers motivation directly contrasts InterpConv with density-sensitive KNN graph construction. The normalization argument in Section 3.2 is a mean-field argument, not a proof of invariance for finite, non-uniform point clouds; the hard support cutoff creates a concrete mechanism by which density changes the computed features. The reader identified exactly this assumption as the weakest, and the paper indeed never stress-tests it in Section 4. I do not see an internal inconsistency or a reason to reject the method: the benchmark numbers are plausible and the operation is clearly specified. However, without a density-resampling experiment, the conditional verdict is appropriate because the advertised property and the small SOTA margins both need further support. I therefore leave the readers verdict unchanged.","tokens_in":14197,"tokens_out":12213,"duration_ms":136987,"concrete_test":"Freeze a trained first-layer InterpConv from the ModelNet40 classifier. Sample the same set of test CAD meshes at 4096 points (dense) and at 512 points (sparse), with output coordinates fixed to the sparse point locations, and compute the per-channel mean squared error between the dense and sparse feature maps. If the relative error is large (e.g., >10%) or scales with the density ratio, the normalization does not deliver practically meaningful sparsity invariance. A complementary end-to-end check would be to evaluate the same trained 1024-point classifier on test clouds randomly subsampled to 512 and 256 points and report accuracy per density; a sharp drop would confirm the network as a whole is density-sensitive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline property of sparsity invariance is established only by the normalization term in Eqs. (7)-(8), not by analysis or experiment. Under i.i.d. thinning of the same sampling distribution, the expectation of Eq. (8) is density independent, so the claim is plausible. But the operation actually aggregates features from points inside a hard-support neighborhood: the 3σ ball for Gaussian interpolation (Eq. (6)) or the enclosing kernel cell for trilinear interpolation (Eq. (5)). For a finite point cloud, the set of points inside that support is a random set whose size and spatial distribution depend on local density. Normalizing by N or Σt turns the sum into a weighted average, but the realized weighted average still depends on which points happen to fall inside the support, and its variance grows as density decreases. In sufficiently sparse regions, the support can even be empty, making the normalization undefined; no fallback is described. The network also compounds the issue: after each 3×3×3 InterpConv layer, output points are downsampled, so later layers operate on progressively sparser clouds. Section 4 contains no experiment that resamples input density, so the claim that InterpConv directly handles irregular inputs better than KNN-based graphs is not empirically supported. This does not invalidate the reported accuracies, but the method's central advantage over prior graph approaches is not yet backed by evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InterpConv, a convolution-like operation for 3D point clouds. Instead of rasterizing points into voxels or building KNN graphs, InterpConv places a regular grid of discrete kernel-weight vectors in 3D space and maps nearby input point features to those weights using either trilinear or Gaussian interpolation. A normalization term divides by the number of neighboring points or by the sum of interpolation weights, which the authors claim makes the operation permutation and sparsity invariant. The paper then builds classification and segmentation networks around InterpConv layers and reports state-of-the-art or competitive results on ModelNet40 (93.0% accuracy), ShapeNet Parts (86.3% instance mIoU), and S3DIS (66.7% mIoU), together with ablations of kernel size, kernel length, interpolation function, normalization, parameter count, and runtime.","tokens_in":14451,"tokens_out":8963,"duration_ms":93535,"significance":"If the claims are borne out, InterpConv is a simple and appealing alternative to graph-based and voxel-based point cloud convolutions: it avoids per-edge MLPs, has a clear formulation in Eq. (3), and the ablations in Section 4.4 isolate the roles of kernel size, kernel length, interpolation, and normalization. The authors also provide parameter-count and inference-time comparisons, which are useful for practitioners. However, the headline property of sparsity invariance is asserted rather than demonstrated, the benchmark margins over strong baselines are small, and no error bars or multiple-run statistics are reported. These issues bear directly on the central claims, so the paper needs additional evidence before it can be accepted as stated.","major_comments":[{"comment":"The claim that InterpConv is sparsity invariant is not supported by the formulation or by experiment. Eqs. (7) and (8) normalize an accumulated sum, but the sum runs over the random set of points inside a hard-cutoff neighborhood: the 3σ ball of Eq. (6) or the kernel cell of Eq. (5). For finite point clouds this set depends on local density; in sufficiently sparse regions it can be empty, and the paper does not specify a fallback for the N=0 or Σt_i=0 case. Normalization by N or Σt_i converts the aggregation into a weighted average, but the realized weighted average still depends on which points fall inside the support, and its variance grows as density decreases. Section 3.3 further downsamples points after each 3×3×3 InterpConv in the segmentation network, so later layers operate on progressively sparser clouds. Section 4 contains no experiment that resamples or thins the input density, so the claimed advantage over KNN-based graph construction is not empirically established. I ask the authors to prove or qualify the invariance under a precise sampling model and to add a density-robustness experiment, for example uniform thinning and nonuniform density resampling.","section":"Section 3.2, Eqs. (7)-(8); Section 4"},{"comment":"All benchmark numbers are reported as single runs without error bars or significance statements. The margins over the strongest baselines are small on ModelNet40 (0.8% over DGCNN/PointCNN in Table 1) and ShapeNet Parts (0.6% over PointConv in Table 2), and the S3DIS result is a single 6-fold mean with no fold-wise variance. Given typical run-to-run variation in point-cloud networks, these differences could be within noise. Please report mean and standard deviation over multiple runs, or at least fold-wise S3DIS results, and state whether the final architecture was selected on a validation split.","section":"Section 4, Tables 1-3 and 4-9"},{"comment":"The interpolation function is chosen per task according to test-set accuracy: Gaussian is used for ModelNet40 and trilinear for ShapeNet Parts, with the selected numbers then reported as the final results. This uses the test set for model selection, which can bias the reported state-of-the-art. The choice should be made on a held-out validation split, or the paper should fix a single selection rule and report the accuracy of both interpolation functions on all tasks under that rule.","section":"Section 4.4, Table 6"},{"comment":"In the classification network the Gaussian bandwidth is fixed to 3σ=0.1, while kernel lengths l range up to 0.8 in Table 5. Because Eq. (6) truncates weights beyond 3σ, points farther than 0.1 from every kernel-weight coordinate receive zero weight; for large l, a substantial fraction of points in the kernel volume may be ignored and some kernel-weight coordinates may have very few or no neighboring points. The paper does not explain how the normalization denominators in Eqs. (7)-(8) are handled in this case, nor whether σ is scaled with l. The effective receptive field is therefore not simply controlled by l. Please clarify the relationship between σ and l, or scale the support with the grid spacing.","section":"Section 4.1, Eq. (6), Table 5"}],"minor_comments":[{"comment":"The pseudocode appears to normalize the accumulated feature inside the loop over neighboring points (\"fi ← fi + t fp; fi ← Normalize(fi)\"), which is inconsistent with Eq. (3). Normalization should occur after the full sum over the neighborhood is accumulated.","section":"Algorithm 1, lines 7-8"},{"comment":"The text says that the input point and kernel weights are normalized into a unit-length cube, but the scaling steps are not written out. Please define the normalization explicitly before presenting the simplified trilinear formula.","section":"Eq. (5)"},{"comment":"Figure 4 shows a qualitative comparison with PointNet++, but PointNet++ is not included in Table 3; either add it to the table or adjust the caption and text.","section":"Section 4.3, Table 3"},{"comment":"The kernel-length triples (e.g., \"0.05-0.1-0.2\") are not explicitly mapped to the three branches of each PointInception module; please make this mapping clear in the table or text.","section":"Section 4.4, Table 5"},{"comment":"The sentence \"We further sample 1,024 points for training and testing\" is ambiguous about whether this is a random subsample or a fixed choice; please specify.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision or 3D learning venue and the proposed operator is conceptually clean. The main risk is that the central property of sparsity invariance is not empirically validated and the reported gains over strong baselines are small without error bars. If the authors add density-robustness experiments, multiple-run statistics, and fix the test-set selection issue, the paper could be acceptable. I do not see grounds for rejection, but the current version should not be accepted without those additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on InterpConv. The core idea is real: instead of learning a continuous kernel with an MLP (PointConv) or a parameterized polynomial (SpiderCNN), the paper fixes a small set of discrete 3D kernel weights and uses an interpolation function (trilinear or Gaussian) to pull point features onto those weight coordinates. That's a clean, sensible design, and the paper presents it clearly. Eq. (3) is simple, and the ablations in Sec. 4.4 are genuinely useful—they isolate kernel size, kernel length, interpolation choice, and normalization, and the conclusions are consistent. The citation pattern is fair: the related work accurately positions this against DGCNN, PointCNN, PointConv, SpiderCNN, and the voxel-based methods. On the benchmark side, the numbers are competitive: 93.0% on ModelNet40, 86.3% instance mIoU on ShapeNet Parts, and 66.7% mIoU on S3DIS, all without normals as input and with inference time well below DGCNN.\n\nThe soft spots are mostly about evidence quality, not about the core construction. The strongest issue is the sparsity-invariance claim. It's asserted in the abstract and Sec. 3.2, and it's motivated by the normalization term, but there is no experiment that resamples point clouds at different densities or that tests behavior on sparse regions. The stress-test note is right: the normalization only makes the aggregation a weighted average over a hard-cutoff neighborhood (the 3σ ball or a kernel cell). In expectation over i.i.d. thinning the result is density-independent, but for finite samples the realized average depends on which points fall inside that support, and in very sparse regions the support can be empty, with no fallback described. So the claim is plausible but unverified. That's a gap, not a fatal flaw.\n\nAlso, the reported margins over prior work are small—0.3% to 0.8%—and there are no error bars or multiple-run statistics. That makes the 'state-of-the-art' claim fragile, though plausible. And there's a minor test-set tuning issue: Table 6 selects the interpolation function per task based on test accuracy, which is a form of peeking. It's a small effect, but worth noting. No code is released, which would have helped verify the results.\n\nOverall, this is a legitimate, useful piece of work. The math is straightforward, and the paper does what it claims to within the limitations I mentioned. I'd send it to peer review without hesitation. The method deserves to be compared against, and the invariance claim should be stress-tested in the revision. I'd bring it to a reading group if you want a discussion of what 'invariance' should mean for point cloud convolutions.","headline":"A solid, clearly-written point cloud convolution paper whose core idea (discrete kernels + interpolation) is real, but whose headline sparsity-invariance claim is asserted rather than demonstrated.","tokens_in":14974,"tokens_out":3914,"would_cite":true,"duration_ms":34976,"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":"InterpConv, a convolution that interpolates point features onto discrete kernel weights, sets new accuracy marks on three 3D point cloud benchmarks.","keywords":["interpolated convolution","point cloud classification","semantic segmentation","sparsity invariance","permutation invariance","discrete kernel weights","multi-scale features","3D deep learning"],"falsifier":"Resample the same ModelNet40 shapes at, say, 256 and 8,192 points, run the trained InterpCNN on both without retraining, and compare accuracy and per-point features; a large drop at low density would refute the sparsity-invariance claim.","tokens_in":13962,"feed_emoji":"🧊","tokens_out":4263,"duration_ms":41099,"temperature":0.7,"pith_summary":"The paper introduces InterpConv, a convolution operation that runs directly on irregular 3D point clouds instead of voxel grids or graphs. Its central claim is that a spatially discrete set of kernel-weight vectors, combined with an interpolation function that pulls nearby point features onto those weights, gives point-cloud networks the same inductive bias that ordinary discrete kernels give image CNNs. A per-weight normalization term is meant to make the operation invariant to how densely points are sampled. The authors build classification and segmentation networks around this operation and report state-of-the-art accuracy on ModelNet40 (93.0%), ShapeNet Parts (86.3% instance mIoU), and S3DIS (66.7% mIoU). A sympathetic reader would take away that discrete convolutional kernels can be adapted to unstructured 3D data without giving up the regularity that makes convolution work.","feed_headline":"InterpConv hits 93.0% on ModelNet40 with discrete kernels","feed_subtitle":"A convolution that interpolates point features onto discrete kernel weights also leads part and scene segmentation benchmarks.","key_machinery":"The central object is the InterpConv operator, written as $$F*W(\\hat{p})=\\sum_{p'} \\frac{1}{N_{p'}}\\sum_{p_\\delta} T(p_\\delta,p')\\,F(\\hat{p}+p_\\delta)\\cdot W(p'),$$ where $W(p')$ are discrete $1\\times c$ kernel-weight vectors with fixed 3D coordinates $p'$, $T$ is an interpolation function (trilinear or Gaussian), and $N_{p'}$ is the density normalization term. The interpolation function is what lets discrete weights act on off-grid points; the normalization term is what carries the sparsity-invariance claim; the fixed cube arrangement of $p'$ is what lets standard receptive-field intuitions (kernel size, kernel length) transfer to point clouds.","core_discovery":"On its own terms, the paper's discovery is that a convolution kernel need not be tied to a regular grid to stay discrete. InterpConv places $n^3$ fixed $1\\times c$ weight vectors at coordinates $p'$ of a cube in 3D space, then for each input point computes interpolation weights $T(p_\\delta,p')$ from the point to nearby weight coordinates. Features are accumulated at each weight coordinate and divided by a normalization term $N_{p'}$ that counts neighboring points or sums their weights, which is what makes the aggregation independent of local density. The convolution at any output location is then the dot product of the normalized feature block with the discrete kernel weights. With trilinear or Gaussian interpolation, this operation is permutation invariant and can be evaluated at arbitrary centers, so it consumes raw point clouds directly while keeping standard convolutional machinery.","pith_inferences":["The sparsity-invariance claim is testable but not directly demonstrated: resampling the same shapes at very different densities and comparing features or accuracy would settle it, and the paper reports no such experiment.","Learnable kernel-weight coordinates, listed as future work, could let kernels adapt to anisotropic local geometry and likely improve segmentation further.","The same interpolation mechanism may transfer to 3D object detection and instance segmentation, where point density varies sharply across scenes.","Other interpolation bases, such as linear B-splines, could trade off localization and smoothness differently and are worth comparing to the two proposed functions."],"forward_implications":["The operation consumes irregular point clouds directly, without voxelization, so fine geometry is preserved and dense-grid cost is avoided.","Multi-branch blocks with different kernel lengths capture both fine-grained local structures and global shape context in a single network.","A U-Net-style encoder-decoder built from InterpConv layers reaches state-of-the-art results on object part and indoor scene segmentation.","The per-weight normalization term makes the convolution invariant to local point density, addressing a weakness of graph-based neighbor selection.","Discrete kernels give point-cloud networks an inductive bias similar to image CNNs while using fewer parameters than continuous MLP-based kernels."],"supporting_citations":[{"why":"Provides the ModelNet40 benchmark and the CAD-model point-cloud conversion used for classification.","marker":"[5]"},{"why":"Represents the voxelization approach that InterpConv aims to replace, motivating direct irregular input.","marker":"[24]"},{"why":"Supplies the baseline PointNet architecture and the 1024-point sampling protocol used in evaluation.","marker":"[25]"},{"why":"Provides the PointNet++ hierarchical grouping baseline and the feature-propagation layers used in the decoder.","marker":"[27]"},{"why":"DGCNN is the strongest graph-based classification baseline that InterpCNN outperforms.","marker":"[45]"},{"why":"PointConv exemplifies continuous convolutional kernels on point clouds, the alternative design InterpConv contrasts with.","marker":"[46]"},{"why":"SpiderCNN supplies a parameterized continuous-kernel baseline and motivates the discrete-weight design.","marker":"[49]"},{"why":"Provides the ShapeNet Parts dataset and part-annotation protocol for object part segmentation.","marker":"[50]"},{"why":"Provides the S3DIS indoor scene dataset and 6-fold validation protocol for semantic parsing.","marker":"[1]"}],"fun_headline_variants":["InterpConv sets new SOTA on 3D shape, part, and scene tasks","Interpolate point features onto discrete kernels for conv on clouds","Discrete kernels, no grid: InterpConv understands raw point clouds","Point cloud conv without voxels: InterpConv tops benchmarks","InterpConv: interpolation brings convolution to unordered point sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the operation is unaffected by how densely points are sampled rests on the normalization term alone, yet the choice of which neighboring points contribute still depends on local point density.","fun_headline_variants_meta":{"raw":{"variants":["InterpConv sets new SOTA on 3D shape, part, and scene tasks","Interpolate point features onto discrete kernels for conv on clouds","Discrete kernels, no grid: InterpConv understands raw point clouds","Point cloud conv without voxels: InterpConv tops benchmarks","InterpConv: interpolation brings convolution to unordered point sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1483,"prompt_tokens":903,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":519,"tokens_out":580,"duration_ms":6087,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:39:52.309901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Resample the same ModelNet40 shapes at, say, 256 and 8,192 points, run the trained InterpCNN on both without retraining, and compare accuracy and per-point features; a large drop at low density would refute the sparsity-invariance claim.","supporting_citations":[{"cited_title":"V oxnet: A 3d con- volutional neural network for real-time object recognition","cited_arxiv_id":null,"evidence_quote":"Represents the voxelization approach that InterpConv aims to replace, motivating direct irregular input."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"Provides the PointNet++ hierarchical grouping baseline and the feature-propagation layers used in the decoder."},{"cited_title":"Spidercnn: Deep learning on point sets with parameterized convolutional filters","cited_arxiv_id":null,"evidence_quote":"SpiderCNN supplies a parameterized continuous-kernel baseline and motivates the discrete-weight design."},{"cited_title":"A scalable active framework for re- gion annotation in 3d shape collections","cited_arxiv_id":null,"evidence_quote":"Provides the ShapeNet Parts dataset and part-annotation protocol for object part segmentation."},{"cited_title":"3d semantic parsing of large-scale indoor spaces","cited_arxiv_id":null,"evidence_quote":"Provides the S3DIS indoor scene dataset and 6-fold validation protocol for semantic parsing."}],"review_version":1}