{"id":"fcf3043b-df41-4fcd-882f-e98eda566148","arxiv_id":"2412.03052","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A residual graph network for point cloud classification and segmentation reports 73.47% mean IoU on S3DIS, with no code or complete architecture specification released.","lead":"Point-GR is a graph-based neural network that classifies and segments 3D point clouds while using fewer parameters than similar models. The authors report a state-of-the-art indoor scene segmentation score on S3DIS, but the paper omits code and gives an incomplete architecture description.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Described residual skip connection in Point-GR Transformation (Sec. 2.1, 2.3) adds mismatched feature maps (N×3 or N×64 to N×6) with no projection layer; the architecture is not implementable as written.","rationale":"The reader's weakest assumption identifies exactly this mismatch, and I agree. The concern is not manufactured: Section 2.3 explicitly fixes the transformation output to [N×6], while Section 2.1's residual add has no compatible operand. This makes the method impossible to implement from the paper, which is a correctness risk of the highest order because the transformation is the shared backbone for classification, part segmentation, and scene segmentation. A missing projection layer could theoretically be a small omission, but then the paper must disclose it; without disclosure, the reported parameter counts (Table 4) and the claimed reduction relative to DG-CNN cannot be audited. The S3DIS comparison is also thin, but that is a secondary weakness about external validity; the internal dimension mismatch is more fundamental because it calls into question whether the described network can compute anything at all. The paper has no code or formal verification, so the textual specification is the only evidence; that evidence is self-inconsistent. Therefore the REJECT verdict is appropriate and unchanged.","tokens_in":9583,"tokens_out":5213,"duration_ms":45615,"concrete_test":"Write a minimal PyTorch implementation of the Point-GR Transformation exactly as specified: Conv2d(3,64), BN, ReLU, then Conv2d(64,6), BN, then add the output of the 'previous layer' (test both the N×3 input and the N×64 first-convolution output) and apply ReLU. A forward pass on a random (N,3) tensor will raise a size-mismatch RuntimeError unless an undeclared projection is inserted. If the authors' intended implementation instead includes a 1x1 convolution, they should state it and show that the Table 4 parameter counts are recomputed with that operation; without that, the architecture is not implementable as described.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the residual connection in the Point-GR Transformation, the foundation of all three reported models, is dimensionally inconsistent with the stated tensor shapes. Section 2.1 says 'the output from the previous layer is added to the output of the convolution layer,' and Section 2.3 says the transformation uses two convolutions: 'first the input dimension is transformed into higher dimension 64 and then second convolution layer squeezes dimension into 6, the final output of Point-GR Transformation is [N × 6].' The only direct readings of 'previous layer' are the block input (N×3) or the first convolution's output (N×64), and both are incompatible with the N×6 second-convolution output. No 1x1 projection, padding, or other dimension-matching operation is described anywhere. Because every subsequent FLN block, concatenation, and parameter count in Table 4 depends on this transformation, the published architecture cannot be instantiated from the manuscript, and the reported 73.47% S3DIS mIoU, 92.71% ModelNet accuracy, and 85.2% ShapeNetPart mIoU cannot be derived or reproduced from the written method. This is an internal inconsistency, not a mere absence of code, so it directly undermines the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Point-GR, a graph-based residual network for point cloud classification, part segmentation, and scene segmentation. The method consists of a Point-GR Transformation module that builds k-NN graphs, computes edge features, and applies residual convolutions, followed by a Point-GR Feature Learning Network with three multi-scale graph blocks. The authors report 92.71% accuracy on ModelNet-40, 85.2% mean IoU on ShapeNetPart, and 73.47% mean IoU on S3DIS, which they claim is state of the art, together with reduced parameter counts relative to the DG-CNN baseline. The paper also presents ablations over input point count and neighborhood size k.","tokens_in":9863,"tokens_out":7516,"duration_ms":66106,"significance":"If the reported results are reproducible, the contribution is potentially useful: a lightweight graph residual network that is competitive on three standard benchmarks and reduces parameters relative to a strong graph baseline would be of interest to the point cloud learning community. The ablation studies in Tables 5 and 6 and the parameter comparison in Table 4 are positive features of the manuscript. The strengths are, however, conditional: there is no code, no error bars, no statistical comparison to prior work, and the S3DIS comparison is too sparse to support the 'state-of-the-art' claim. More importantly, the central architectural description is internally inconsistent, so the reported numbers cannot currently be derived from the manuscript as written.","major_comments":[{"comment":"The residual skip connection in the Point-GR Transformation is dimensionally inconsistent. Section 2.1 states that 'the output from the previous layer is added to the output of the convolution layer,' while Section 2.3 defines the transformation as two convolutions that map the input first to 64 channels and then to 6 channels, with final output [N x 6]. The only candidate 'previous layer' outputs are the block input (N x 3) and the first convolution's output (N x 64); both are incompatible with the N x 6 output of the second convolution. No 1x1 projection, padding, or other dimension-matching operation is specified anywhere in the manuscript. Since every downstream FLN block, concatenation, and the parameter count in Table 4 depends on the exact output of this transformation, the architecture cannot be instantiated from the written method, and the reported classification, part-segmentation, and S3DIS results cannot be reproduced or verified.","section":"Section 2.1/2.3"},{"comment":"The claim in the abstract and Section 3.3 that Point-GR achieves state-of-the-art scene segmentation is not supported by Table 3, which compares only PointNet, PointCNN, and DG-CNN. Several established methods (e.g., PointNet++, KPConv, and more recent S3DIS submissions) are absent, and no quantitative comparison to them is provided. A 'state-of-the-art' claim requires comparison with the current best published methods under the same evaluation protocol. The 73.47% mean IoU is therefore a standalone result, not evidence of state-of-the-art performance.","section":"Table 3 / Section 3.3"},{"comment":"The S3DIS input features are described inconsistently. The authors first say each point has '9 attributes (XYZ, RGB, and normalized spatial coordinates)' but two sentences later say the input contains '9 channels x, y, z, RGB color information, intensity, normal vector.' The second list has more than nine components (XYZ, RGB, intensity, and a 3D normal vector sum to ten), so the actual input dimension is unclear. Because the input feature dimension affects both the network architecture and the S3DIS result, this ambiguity is load-bearing for reproducibility.","section":"Section 3.3"},{"comment":"All benchmark results are single-run accuracies with no error bars, standard deviations, seeds, or code release. In addition, the manuscript switches between 'DG-CNN' and 'DC-GNN' (Table 1 lists both DG-CNN [9] at 92.9% and DC-GNN [17] at 93.64%; Table 2 uses DC-GNN; Section 3.1 refers to 'the baseline DG-CNN model'). If DC-GNN is the authors' own prior model, this is not disclosed in the main text, and the comparison is not a standard independent baseline. The small accuracy gaps in Table 1 (0.19-0.93 percentage points) are within typical run-to-run variation for point cloud networks, so without repeated runs the comparative claims are not statistically supported.","section":"Tables 1-3 / Section 3"}],"minor_comments":[{"comment":"The class-wise IoU numbers in the DC-GNN row are run together (e.g., '84.53 78.95 54.28 85.02 95.8 85.38 81.0 92.2 82.9'), making the table very hard to read.","section":"Table 2"},{"comment":"Table 1 cites P2Sequence as reference [20], but reference [20] is titled 'PCT: Point Cloud Transformer'; the citation appears to be mismatched.","section":"Table 1 / References"},{"comment":"The PointNet row in Table 4 contains a single parameter value (0.69M) with no task breakdown, while DG-CNN and Ours have values for Classification, Part-Seg, and S3DIS; please clarify or align the entries.","section":"Table 4"},{"comment":"The claim that residual-based learning mitigates point permutation problems is not justified: permutation invariance in this architecture comes from symmetric aggregation (max pooling) in the graph convolution, not from residual skip connections. Please rephrase or provide a concrete argument.","section":"Abstract / Section 1"},{"comment":"The sentence 'Our model demonstrates average performance with an accuracy of 92.71%, surpassing most of the state-of-the-art networks' is misleading because 92.71% is below several entries in Table 1, including RS-CNN, KPConv, DG-CNN, and DC-GNN.","section":"Section 3.1"},{"comment":"There are numerous typos and grammatical errors, including 'Psuedo' in Algorithm 1, 'Classfication' in Algorithm 1, and 'eliminate the issue' in Section 2.1; the manuscript would benefit from careful proofreading.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The central problem is not the absence of code alone; it is that the described architecture contains a tensor-shape contradiction in the module that is common to all three task networks. This is an internal inconsistency in the method as written, and fixing it would require restating the residual path and re-running all experiments. Combined with the sparse S3DIS comparison and the lack of statistical support, the manuscript is below the publication bar in its current form. The idea of a residual graph transformation is plausible, and a resubmission with code, a clarified residual projection, and a properly populated S3DIS comparison could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on arXiv:2412.03052. The paper proposes a lightweight graph residual network for point cloud tasks. The core idea — a residual block that transforms raw point coordinates to a compact feature before feeding a multi-scale feature learning network — is a legitimate if modest extension of DGCNN and the authors' earlier DC-GNN. The parameter-efficiency comparison in Table 4 and the ablations over k and point count are genuinely useful. If the S3DIS number (73.47 mIoU) held up, it would be a meaningful data point.\n\nBut the paper has a load-bearing problem: the residual connection in the Point-GR Transformation is dimensionally undefined. Section 2.3 says the transformation's final output is [N×6] after two convs (input to 64, then to 6). Section 2.1 says a skip connection adds 'the output from the previous layer' to the convolution output. The previous layer can be the block input (N×3) or the first conv output (N×64), and neither matches N×6. No projection layer is described anywhere. Since every following FLN block, concatenation, and the parameter counts in Table 4 depend on this transformation, the architecture cannot be instantiated from the manuscript. That's an internal inconsistency, not just missing code.\n\nOther issues are significant but secondary. The S3DIS comparison in Table 3 omits essentially all modern baselines, so calling 73.47 'state-of-the-art' is unsupported. The input feature list in Section 3.3 is inconsistent (9 channels vs. XYZ, RGB, intensity, normal = 10). No error bars or code are provided. And the claim that residual learning solves permutation invariance is conceptually off — permutation invariance comes from the symmetric max aggregation, not the skip connection. The heavy reliance on the authors' own DC-GNN as baseline is not itself a flaw, but it deserves more disclosure in the main text.\n\nWhat's genuinely good: the paper is clearly written in places, the ablation study over k and point count is reasonable, and the parameter reduction compared to DGCNN is a real consideration. The authors are engaged with the literature, and the idea of a residual transformation block is worth exploring.\n\nWho is this for? Someone working on lightweight graph point cloud models might find the initial idea interesting, but as written, there's nothing to build on. A serious referee could straighten this out if the authors clarify the residual path (e.g., add a 1×1 conv) and release code. I'd send it to peer review rather than desk reject, but with the expectation that the authors must fix the architecture description and provide code or strong justification for the numbers. As it stands, my recommendation is: don't trust the numbers until the architecture closes.","headline":"The residual graph idea is plausible, but the architecture as written doesn't close dimensionally and the S3DIS claim is unverifiable.","tokens_in":10380,"tokens_out":5180,"would_cite":false,"duration_ms":46478,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A residual graph network for raw point clouds reaches 73.47% mean IoU on indoor scene segmentation while using fewer parameters than graph-based baselines.","keywords":["point cloud","3D object classification","part segmentation","scene segmentation","graph neural network","residual learning","dynamic graph"],"falsifier":"Implement the Point-GR Transformation exactly as written and run a forward pass on ModelNet-40: if the residual addition has no specified channel projection, the shapes will not align and the network will not train; alternatively, run the same pipeline with and without an inserted 1x1 projection and check whether the 73.47% S3DIS and 92.71% ModelNet-40 numbers reproduce under the stated hyperparameters.","tokens_in":9367,"feed_emoji":"","tokens_out":6261,"duration_ms":55082,"temperature":0.7,"pith_summary":"The paper proposes Point-GR, a graph-based deep network that takes unordered point clouds and first transforms them through a residual embedding block before learning hierarchical features. Its central claim is that this residual transformation preserves local geometry while shrinking the parameter count relative to comparable graph networks, and that the resulting model achieves the best reported mean IoU (73.47%) on the S3DIS indoor scene segmentation benchmark. On ModelNet-40 it reports 92.71% classification accuracy and on ShapeNetPart 85.2% mIoU, enough to match the strongest existing graph methods. A sympathetic reader would take the contribution as evidence that residual graph embeddings offer a lightweight backbone that transfers across classification, part segmentation, and scene segmentation.","feed_headline":"Point-GR lifts S3DIS scene segmentation to 73.47% mean IoU","feed_subtitle":"A residual graph network matches top 3D classifiers with fewer parameters than graph baselines.","key_machinery":"The central object is the Point-GR Transformation block: graph construction via k-nearest neighbors, edge features $d_{pe} = [d_p, d_e]$ with $d_p = P_i$ and $d_e = P_i - P_j$, followed by a Point Residual Embedding that interleaves convolution, batch normalization, ReLU, a second convolution, and a residual skip connection before max selection over the local k-neighborhood. This block is the reusable front end for all three tasks, and it is paired with a Feature Learning Network that stacks Multi-Scale Graph Construction and Point Embedding modules at channel depths $\\{6,64,128,256\\}$ and concatenates per-stage skip features. The residual path is what the authors credit for stable training and parameter savings.","core_discovery":"Point-GR is built from a Point-GR Transformation module that constructs a k-NN graph over the input points, forms edge features by concatenating each point's own coordinates with the difference to its neighbors ($d_p$ and $d_e$), and then runs two convolutions with a residual skip connection between the previous layer output and the convolution output. The resulting point residual features are passed through repeated Feature Learning blocks that build multi-scale graphs at progressive channel dimensions {6,64,128,256}, and skip connections from each stage are concatenated before a final convolutional and fully connected head. The paper claims this design yields 92.71% overall accuracy on ModelNet-40, 85.2% mean IoU on ShapeNetPart, and 73.47% mean IoU with 82.79% overall IoU on S3DIS, together with fewer trainable parameters than the graph baseline on classification (1.80M vs 2.61M) and part segmentation (1.04M vs 1.45M). The residual transformation is the load-bearing novelty that the authors say mitigates point-permutation sensitivity and vanishing gradients while keeping the network lightweight.","pith_inferences":["The paper does not specify how the residual skip connection aligns channels when the previous layer's feature count differs from the 6-channel output of the transformation; the described addition may require an implicit 1x1 projection or padding, and the architecture is only fully reproducible once that step is pinned down.","If the 73.47% S3DIS number holds under a standard 6-fold protocol, the largest gap to prior graph baselines may come more from training setup (block size, 9-channel input, optimizer schedule) than from the residual mechanism itself; an ablation that swaps only the residual path on and off would separate these contributions.","The same residual transformation could be tested on sparser and noisier LiDAR scans; the k-NN graph construction plus residual embedding is a natural candidate for outdoor point clouds where density varies sharply."],"forward_implications":["A single residual graph backbone can serve classification, part segmentation, and indoor scene segmentation, so the method should transfer to other point-cloud tasks such as object detection and instance segmentation without redesign.","The reported parameter reductions (1.80M vs 2.61M for classification; 1.04M vs 1.45M for part segmentation) imply lower memory and faster inference for the same accuracy class, which matters on embedded and LiDAR platforms.","The S3DIS result, if taken at face value, suggests a substantial improvement over graph-based baselines on scene-level semantic segmentation, moving the graph approach closer to the top of the benchmark.","The ablation sensitivity to k (peak at k=20, degradation at k=40) gives a practical recipe: the neighborhood size should be tuned rather than defaulted to large values."],"supporting_citations":[{"why":"Supplies the point-cloud baseline and the permutation-invariance formulation that Point-GR extends, and defines the S3DIS preprocessing standard.","marker":"[5]"},{"why":"Provides the dynamic graph convolution baseline that Point-GR compares against on accuracy and parameter counts.","marker":"[9]"},{"why":"Earlier graph network from the same group that Point-GR builds on and compares with on classification and part segmentation.","marker":"[17]"},{"why":"Defines the ModelNet-40 benchmark and its standard train/test split used for classification evaluation.","marker":"[4]"},{"why":"Provides the ShapeNet-Part dataset and its part-segmentation evaluation protocol.","marker":"[28]"},{"why":"Supplies a hierarchical point-based method that motivates the multi-scale Feature Learning blocks.","marker":"[6]"}],"fun_headline_variants":["Residual graph net: 73.47% S3DIS IoU, fewer params","Point-GR cuts graph-net parameters, hits 73.47% S3DIS","Graph residual network: 92.71% ModelNet40, 73.47% S3DIS","Point-GR: residual edges, fewer weights, top scene segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the described residual block is actually constructible: the skip connection must add feature maps whose channel dimensions are not stated to match, since the final transformation output is [N x 6], while the layer feeding the residual is the raw input or a 64-channel feature map.","fun_headline_variants_meta":{"raw":{"variants":["Residual graph net: 73.47% S3DIS IoU, fewer params","Point-GR cuts graph-net parameters, hits 73.47% S3DIS","Graph residual network: 92.71% ModelNet40, 73.47% S3DIS","Point-GR: residual edges, fewer weights, top scene segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1429,"prompt_tokens":930,"completion_tokens":499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":546,"tokens_out":499,"duration_ms":4461,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:49:26.607611+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the Point-GR Transformation exactly as written and run a forward pass on ModelNet-40: if the residual addition has no specified channel projection, the shapes will not align and the network will not train; alternatively, run the same pipeline with and without an inserted 1x1 projection and check whether the 73.47% S3DIS and 92.71% ModelNet-40 numbers reproduce under the stated hyperparameters.","supporting_citations":[{"cited_title":"Pointnet: Deep learning on point sets for 3d classification and segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the point-cloud baseline and the permutation-invariance formulation that Point-GR extends, and defines the S3DIS preprocessing standard."},{"cited_title":"Dynamic graph cnn for learning on point clouds","cited_arxiv_id":null,"evidence_quote":"Provides the dynamic graph convolution baseline that Point-GR compares against on accuracy and parameter counts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier graph network from the same group that Point-GR builds on and compares with on classification and part segmentation."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"Defines the ModelNet-40 benchmark and its standard train/test split used for classification evaluation."},{"cited_title":"A scal- able active framework for region annotation in 3d shape collections.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the ShapeNet-Part dataset and its part-segmentation evaluation protocol."}],"review_version":1}