{"id":"00840e23-b4c6-4b4f-9bc2-35a891b7e94d","arxiv_id":"2508.19003","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RoofSeg, an end-to-end transformer with edge-aware and geometric losses, reports state-of-the-art roof plane segmentation on three benchmarks.","lead":"This paper presents RoofSeg, a transformer network that segments roof planes directly from airborne LiDAR point clouds without clustering post-processing, using edge-aware refinement and geometric losses. It reports the highest scores on three roof segmentation benchmarks, though code and a recent transformer baseline are missing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing baseline: SPPSFormer (Zeng et al. 2025), cited as a roof-plane transformer, is absent from Table 1; without it the SOTA claim is unverified.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict, but for a different primary reason. The reader's weakest_assumption focuses on the non-differentiable hard binarization in EAMM and the adaptive weighting loss. That is a legitimate concern about the 'truly end-to-end' claim and about whether EAMM is trained as described, but it does not directly threaten the reported performance numbers: even if some operations are non-differentiable and treated as constants, the model can still be trained and the mCov values can still be valid. The most load-bearing concern is the omitted SPPSFormer baseline. The paper's own literature review identifies SPPSFormer as a competing transformer for roof plane segmentation, yet Table 1 excludes it. Since the central claim is state-of-the-art performance, a missing directly relevant baseline makes the claim unsubstantiated. This is easily testable and has a clear binary outcome. The paper otherwise shows thorough ablations, consistent improvements, and careful benchmark descriptions, so I would not reject it outright; conditional acceptance pending the SPPSFormer comparison is appropriate. Thus the verdict remains CONDITIONAL (UNCHANGED).","tokens_in":20069,"tokens_out":4227,"duration_ms":39254,"concrete_test":"Run SPPSFormer (Zeng et al., 2025) on the same RoofNTNU, Roofpc3D, and Building3D test splits used in Table 1, using the same metrics (mCov, mWCov, mPrec, mRec). If the authors cannot obtain the official code, reimplement the method from the arXiv paper. Report results over at least three random seeds to also address the lack of error bars. If SPPSFormer's mCov is within 1-2 points of RoofSeg on any of the three benchmarks, the claim of 'significantly outperforming' is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RoofSeg 'significantly outperforms the current competitive approaches' and achieves state-of-the-art results on three benchmarks. Table 1 compares against Region Growing, RANSAC, GoCoPP, PointGroup, Mask3D, and DeepRoofPlane, but omits SPPSFormer (Zeng et al., 2025), an arXiv preprint from May 2025 that also targets roof plane instance segmentation with a transformer architecture. The paper explicitly cites SPPSFormer in Section 2.2, describing it as a superpoint-based transformer that uses post-processing, yet never evaluates against it. Because SPPSFormer is a directly relevant, near-contemporary strong baseline, its omission means the reported 3-4 point mCov advantage over DeepRoofPlane does not establish SOTA. If SPPSFormer performs comparably or better, the core claim collapses. This is a concrete, falsifiable gap in the experimental evidence, independent of the differentiability concern about EAMM, which affects the 'truly end-to-end' narrative but not the validity of the reported numbers.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents RoofSeg, a query-based transformer for end-to-end roof plane instance segmentation from airborne LiDAR point clouds. It combines a PointNet++ encoder with attention-based feature propagation, multiple query refinement decoders, an Edge-Aware Mask Module (EAMM) that augments features at predicted edge points with point-to-plane distances, and a loss function with adaptive point weighting plus a plane geometric term. Experiments are reported on RoofNTNU, Roofpc3D, and Building3D; RoofSeg is claimed to outperform Region Growing, RANSAC, GoCoPP, PointGroup, Mask3D, and DeepRoofPlane on mCov/mWCov/mPrec/mRec. Ablations and efficiency analyses support the main design choices.","tokens_in":20351,"tokens_out":6819,"duration_ms":59972,"significance":"If the results are reproducible and the comparison is complete, this is a useful advance for LoD2/LoD3 building reconstruction. The paper's strengths include consistent gains across three datasets, a systematic ablation of EAMM and the loss components, and an efficiency analysis of decoder count and point resolution. The central claims, however, rest on a comparison set that omits a directly relevant transformer baseline, on an unclear treatment of non-differentiable operations inside the claimed end-to-end training loop, and on single-run point estimates without variance. These gaps do not by themselves invalidate the method, but they need to be resolved before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"SPPSFormer (Zeng et al., 2025) is described in Section 2.2 as a superpoint-based transformer for roof plane instance segmentation, yet it does not appear in the quantitative comparison of Table 1. This is a near-contemporary, directly comparable transformer baseline for the exact task. Without it, the claim of \"new state-of-the-art\" is not established. Please add the comparison or explicitly justify why it is excluded.","section":"Section 2.2 / Table 1"},{"comment":"The paper claims truly end-to-end training. However, Eq. (3) binarizes predicted masks at a 0.5 threshold; the edge mask is also binarized; EAMM extracts edge/interior points, fits a plane with PCA, and computes point-to-plane distances; and Algorithm 1 identifies outliers via a KNN majority rule. These are non-differentiable operations, and the manuscript does not state whether gradients are detached, softened, or approximated. Without this information, the end-to-end claim and the attribution of the reported gains to EAMM are not established. Please specify the gradient flow through these components or revise the claim.","section":"Section 3.3 / Eq. (3) / Algorithm 1"},{"comment":"All metrics are reported as single-run point estimates. Some differences are small (e.g., Table 5 shows mCov differences under 0.5 percentage points; Table 9 shows differences around 0.0005-0.001 between multi-scale and full-resolution features). The phrase \"significantly outperforms\" is therefore not supported by the presented evidence. Please report mean and standard deviation over at least three independent runs, or otherwise quantify variability.","section":"Tables 1-9"},{"comment":"Roofpc3D was generated by Li et al. (2024) and the Building3D plane labels were manually annotated by Li et al. (2024); the evaluation metrics are also defined in that paper. Since the authors are from the same group, there is a risk of annotation or evaluation bias. Please state the relationship explicitly, describe the labeling protocol, and, if possible, include a fully external benchmark beyond RoofNTNU.","section":"Section 4.2"}],"minor_comments":[{"comment":"In the weighted Dice loss, the denominator appears as w_j(Σ a_j + Σ lgt_j) + ε with w_j outside the sums. This is dimensionally inconsistent; the intended form is likely 2Σ w_j a_j lgt_j / (Σ w_j a_j + Σ w_j lgt_j). Please correct.","section":"Eq. (9)"},{"comment":"Line 10 compares \"Nmis\" which is not defined; it should presumably be N_j^dif. Also the loop in line 3 says \"for j = 0 to N\" but the text indexing is 1-based elsewhere.","section":"Algorithm 1"},{"comment":"Typo: \"aiborne LiDAR\" should be \"airborne LiDAR\".","section":"Abstract"},{"comment":"Typos: \"sampling radio\" should be \"sampling ratio\"; \"output demision\" should be \"output dimension\"; \"dropout threshold\" is likely \"dropout rate\".","section":"Section 4.1"},{"comment":"Caption reads \"BAMM\" in the third row description; this should be \"EAMM\".","section":"Figure 10 caption"},{"comment":"Typo: \"The input point clouds can be donated\" should be \"denoted\".","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the method is reasonable. The two most important risks are the omission of the closely related SPPSFormer baseline and the lack of clarity about how non-differentiable operations inside EAMM and the adaptive weighting loss are handled during backpropagation. Editors may also wish to require a clear statement about the authors' relationship to the Roofpc3D and Building3D benchmarks, since those were generated/annotated in prior work by the same group. The promise to release code is welcome and should be honored."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RoofSeg is a well-engineered query-transformer for roof plane segmentation. It adapts the Mask3D/SPFormer line to a specialized geospatial problem, adds an edge-aware module that uses point-to-plane distances, plus adaptive weighting and geometric losses. The ablations are genuinely thorough: EAMM, each loss term, backbone, query count, neighbor count, decoder depth, multi-scale vs full-res features. That is more systematic than a lot of what lands in this subfield. On three benchmarks the method beats the listed baselines by 3–4 points in mCov, and the qualitative results look clean.\n\nSoft spots, in order of importance. The paper cites SPPSFormer (Zeng et al. 2025) as a transformer-based roof-plane competitor but never benchmarks it. Given the intro claims “significantly outperforming current competitive approaches,” that omission is a gap; it needs to be in Table 1 or justified. Second, there are no error bars or repeated runs. The margins are large enough that I suspect they hold, but with single seeds you can't rule out lucky runs on the self-created benchmarks. Third, the benchmarks themselves are partly in-house: Roofpc3D was generated by the same group and Building3D labels were annotated in their prior paper. RoofNTNU is external, which helps. Fourth, the “truly end-to-end” claim is underspecified: Eq. (3) binarizes masks, the edge mask is binarized, and Algorithm 1 makes hard KNN decisions inside the training loop. The paper doesn't state whether gradients are detached or approximated. Likely they are detached or straight-through, but it matters for the EAMM and weighting loss to actually be learnable as described.\n\nNone of this sinks the paper. The core architecture is sound, the ablations support the contributions, and the missing pieces are all addressable in revision.\n\nRecommendation: send to peer review. A good reviewer will ask for the SPPSFormer comparison, variance estimates or a note on why they're omitted, and a precise description of gradient flow through the hard decisions. With those, this is a useful reference. I'd cite it if I worked on 3D building reconstruction.","headline":"RoofSeg is a serious end-to-end transformer for roof-plane segmentation with thorough ablations; the SOTA claim needs a missing baseline and a differentiability clarification, both fixable.","tokens_in":20836,"tokens_out":2966,"would_cite":true,"duration_ms":27151,"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":"RoofSeg claims that roof planes in airborne LiDAR point clouds can be segmented end to end by a transformer using learnable plane queries, with edge-aware refinement and geometric losses, reporting roughly 3-4 point improvements over the pr","keywords":["roof plane segmentation","LiDAR point clouds","transformer","edge-aware mask module","3D building reconstruction","instance segmentation","point cloud segmentation"],"falsifier":"Measure the gradient of the total loss L with respect to the pre-binarization plane-mask logits (the sigmoid(A) values before the 0.5 threshold in Eq. 3) and to the fused edge features E_fuse in EAMM. If RoofSeg trains end-to-end as claimed, these gradients are nonzero and the tangent-distance vector changes training; if the binarization and KNN outlier selection block the gradient, the edge-aware module cannot be learning from the geometric cue, and the ablation gains would have to come from another mechanism. A simpler discriminator: replace the tangent-distance vector in EAMM with random va","tokens_in":19954,"feed_emoji":"🏠","tokens_out":6457,"duration_ms":57037,"temperature":0.7,"pith_summary":"RoofSeg argues that roof plane segmentation from airborne LiDAR can be done truly end to end: a transformer with learnable plane queries directly predicts plane instance masks, eliminating hand-tuned clustering and post-processing. To fix poor edge discrimination, it adds an Edge-Aware Mask Module that injects point-to-plane distances at predicted edge points, plus an adaptive-weighting mask loss and a plane geometric loss that reduce misclassified points and enforce planarity. On RoofNTNU, Roofpc3D, and Building3D, it reports mean coverage of 0.9589, 0.9601, and 0.9374, roughly 3-4 percentage points above the strongest prior approach. If correct, this would make roof plane extraction for 3D building reconstruction less sensitive to geometric clustering hyperparameters.","feed_headline":"RoofSeg reads rooftop LiDAR straight to plane labels","feed_subtitle":"Edge-aware transformer beats clustering pipelines by 3-4 points on three roof benchmarks.","key_machinery":"The load-bearing mechanism is the set of learnable plane queries interacting with multi-scale point features through hierarchical cross-attention (the Query Refinement Decoders), followed by the Edge-Aware Mask Module (EAMM). EAMM augments the features of predicted edge points with the tangent distance from each edge point to a PCA-fitted plane of the predicted mask, then runs a self-attention layer over the merged edge and non-edge features to produce refined masks. Two loss terms carry the training signal: an adaptive-weighting BCE+Dice mask loss that suppresses misclassified points, and a plane geometric loss that minimizes distance from in-plane points to the fitted plane. The claim of '","core_discovery":"The central claim is that roof planes in airborne LiDAR point clouds can be segmented in a single network forward pass by a query-based transformer, without iterative region growing, RANSAC, or clustering. A fixed set of learnable plane queries is refined through hierarchical cross-attention over multi-scale PointNet++ features; each query then predicts an initial plane mask. The Edge-Aware Mask Module (EAMM) refines these masks by extracting predicted edge points, fitting a PCA plane to the predicted interior points, and fusing the edge points' tangent distances to that plane back into the point features. An adaptive weighting scheme in the mask loss lowers the influence of points whose lab","pith_inferences":["If the end-to-end claim holds, the same recipe of learnable queries plus explicit point-to-plane residuals could transfer to other planar primitive tasks, such as indoor wall and floor segmentation, where edge precision is the bottleneck.","The non-differentiable binarization and KNN selection inside EAMM could be replaced with differentiable soft assignments (e.g., straight-through estimators or weighted least-squares plane fitting), which would make the 'truly end-to-end' claim unambiguous and potentially improve the edge cue.","A testable prediction from the ablations is that EAMM's gain scales with the density of edge points: on roofs with few adjacent planes, the module should contribute less than the reported 4-5 points in mCov, while on highly fragmented roofs it should contribute more.","The reported insensitivity to the number of nearest neighbors in adaptive weighting suggests the mechanism acts as a broad regularizer rather than a finely tuned outlier filter, which may simplify transfer to datasets with different point densities."],"forward_implications":["Roof plane segmentation for LoD2/LoD3 building reconstruction could become a single network forward pass, removing the hyperparameter tuning burden of clustering and region growing.","Edge accuracy should improve on complex roofs because point-to-plane distance gives the network an explicit geometric cue where features are usually least discriminative.","The plane geometric loss should make predicted segments lie closer to true planar surfaces, simplifying downstream model fitting and boundary extraction.","On all three benchmarks, RoofSeg outperforms both traditional pipelines and the strongest deep baseline, suggesting a new state of the art for this task.","The adaptive weighting and geometric loss, being largely insensitive to neighbor-count settings, point toward a stable training recipe that could transfer to other planar instance segmentation problems."],"supporting_citations":[{"why":"Supplies the query-based 3D mask transformer design, the base mask loss (BCE+Dice), and the bipartite-matching cost that RoofSeg adapts.","marker":"Schult et al., 2023"},{"why":"Provides the PointNet++ backbone with set abstraction and feature propagation that RoofSeg uses for multi-scale point features.","marker":"Qi et al., 2017"},{"why":"Supplies the Fourier positional encoding used to embed plane query points for unordered point clouds.","marker":"Misra et al., 2021"},{"why":"Introduces the edge-aware transformer idea for 3D instance segmentation that RoofSeg extends with explicit geometric priors.","marker":"Roh et al., 2024"},{"why":"The strongest baseline (DeepRoofPlane) and the source of the Roofpc3D benchmark and Building3D plane labels used for evaluation.","marker":"Li et al., 2024"},{"why":"Provides the RoofNTNU benchmark and the multi-task pointwise network baseline RoofSeg compares against.","marker":"Zhang and Fan, 2022"},{"why":"Establishes the end-to-end masked-attention mask transformer paradigm that motivates the query-based plane mask prediction.","marker":"Cheng et al., 2022"}],"fun_headline_variants":["RoofSeg: one pass from LiDAR points to roof plane masks","Edge-aware transformer segments roof planes from LiDAR in one shot","One pass: transformer turns LiDAR roof points into plane masks","RoofSeg's edge-aware transformer lifts roof plane segmentation accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The hard threshold at 0.5 in Eq. (3), the binarized edge mask, and the KNN majority-rule outlier selection in Algorithm 1 are treated as trainable parts of an end-to-end network, but the paper never specifies how gradients pass through them; if they do not, the claim of truly end-to-end edge-aware training is not established.","fun_headline_variants_meta":{"raw":{"variants":["RoofSeg: one pass from LiDAR points to roof plane masks","Edge-aware transformer segments roof planes from LiDAR in one shot","One pass: transformer turns LiDAR roof points into plane masks","RoofSeg's edge-aware transformer lifts roof plane segmentation accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001385,"raw_usage":{"total_tokens":5480,"prompt_tokens":818,"completion_tokens":4662,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":4589}},"tokens_in":562,"tokens_out":4662,"duration_ms":30365,"temperature":1.0,"reasoning_tokens":4589,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:02:09.813538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the gradient of the total loss L with respect to the pre-binarization plane-mask logits (the sigmoid(A) values before the 0.5 threshold in Eq. 3) and to the fused edge features E_fuse in EAMM. If RoofSeg trains end-to-end as claimed, these gradients are nonzero and the tangent-distance vector changes training; if the binarization and KNN outlier selection block the gradient, the edge-aware module cannot be learning from the geometric cue, and the ablation gains would have to come from another mechanism. A simpler discriminator: replace the tangent-distance vector in EAMM with random va","supporting_citations":[{"cited_title":", author Engelmann, F","cited_arxiv_id":null,"evidence_quote":"Supplies the query-based 3D mask transformer design, the base mask loss (BCE+Dice), and the bipartite-matching cost that RoofSeg adapts."},{"cited_title":", author Girdhar, R","cited_arxiv_id":null,"evidence_quote":"Supplies the Fourier positional encoding used to embed plane query points for unordered point clouds."},{"cited_title":", author Jung, H","cited_arxiv_id":null,"evidence_quote":"Introduces the edge-aware transformer idea for 3D instance segmentation that RoofSeg extends with explicit geometric priors."},{"cited_title":", author Fan, H","cited_arxiv_id":null,"evidence_quote":"Provides the RoofNTNU benchmark and the multi-task pointwise network baseline RoofSeg compares against."},{"cited_title":", author Misra, I","cited_arxiv_id":null,"evidence_quote":"Establishes the end-to-end masked-attention mask transformer paradigm that motivates the query-based plane mask prediction."}],"review_version":1}