acceptodds
Under review as a conference paper at ICLR 2027

HASP-Net: Manifold-Guided Point-Cloud Segmentation of High-Aspect-Ratio Scences

Abstract

Point cloud segmentation has advanced substantially in recent years, yet scenes containing high-aspect-ratio (HAR) structures and small attached components remain challenging. In these scenes, elongated structures often extend far beyond local neighborhoods while lying close to geometrically similar objects, making it difficult to simultaneously preserve long-range continuity and avoid feature interference from adjacent structures. Sparse sampling and scene cropping further weaken structural cues, while conventional point-wise metrics may underrepresent errors such as local breaks and missed attachments. To address this gap, we propose HASP-Net, a manifold-guided segmentation framework built on PTv3 for modeling long-range structural continuity and local semantic variation. The Long-range Track Module represents elongated structures as low-dimensional manifolds in 3D space, constructs geometry-consistent support chains, and propagates context bidirectionally along them to capture long-range dependencies while reducing interference from nearby structures. The Geometry-guided Interface Module focuses on regions with complex local manifold geometry and refines their features through geometry-aware point-set reasoning, reliability-gated residual updates, and semantic contrastive supervision. Together, the two modules jointly model long-range manifold structure and local geometric variation for improved segmentation of elongated structures and their attached components. However, existing benchmarks typically focus on specific domains and lack a unified setting for evaluating these challenges. We introduce HAR3D-Seg, a unified benchmark covering transmission, railway, and vascular scenes, with the Semantic Continuity Score (SCS) and Cross-class Adjacency Score (CAS) for structural evaluation. Averaged across the three scene types, HASP achieves mIoU, mAcc, OA, SCS, and CAS, demonstrating strong segmentation performance across diverse scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point-cloud semantic segmentation, high-aspect-ratio structures, manifold-guided geometric learning, 3D segmentation benchmarks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oXVdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies point-cloud semantic segmentation in high-aspect-ratio (HAR) scenes, focusing on long-range semantic inconsistency and class confusion near attached components. Building on PTv3, the authors propose HASP-Net, which employs geometry-consistent ordered support chains for long-range context aggregation, geometry-guided local refinement, and semantic contrastive learning. The authors also introduce HAR3D-Seg, combining newly collected transmission point clouds with public railway and vascular datasets, along with two structural evaluation metrics, SCS and CAS. Experiments across the three scene types report improved segmentation performance, supported by ablation studies and inference-efficiency comparisons.

Critical strengths
  1. The Long-range Track Module organizes local geometric cells into ordered support chains and aggregates context using arc-length and reliability weighting. This design effectively incorporates the geometric prior of elongated structures and provides a meaningful structural inductive bias for HAR segmentation.

  2. The authors collect and annotate large-scale UAV-LiDAR transmission point clouds and combine them with public railway and vascular datasets to construct HAR3D-Seg. The additional SCS and CAS metrics provide useful complementary perspectives on structural continuity and local class discrimination.

Critical weaknesses
  1. While the ordered support chains are tailored to HAR scenes, the framework largely combines established techniques, including geometric propagation, attention-based refinement, gated residual updates, and contrastive learning. Its methodological distinction from existing structure-aware methods remains unclear. Moreover, the ablations do not fully support the intended mechanisms. In Table 6, Long-range alone improves mIoU but not SCS (1.83 to 1.82), while Interface alone reduces CAS (1.00 to 0.95). Interface+CL also achieves higher CAS than the full model (1.08 vs. 1.03). These results leave the individual contributions and interactions of the proposed modules insufficiently explained.

  2. PTv3 uses spherical crops of up to 120,000 points, whereas HASP uses scene-wide samples of up to two million points. These differences may affect both accuracy and efficiency. Notably, HASP(Base) already achieves substantially higher OA than PTv3 (95.50% vs. 78.06%), despite slightly lower mIoU (Table 3). The inference-efficiency comparison also involves different cropping and voting pipelines. Thus, matched-protocol comparisons are needed to better isolate the benefits of the proposed method.

Questions for authors
  1. In Table 6, Long-range alone does not improve SCS, whereas adding it to Interface+CL increases SCS from 1.84 to 1.92. How do the authors explain this interaction? Why does Interface alone reduce CAS, and why does Interface+CL outperform the full model on CAS? Could a CL-only control clarify these effects? Additionally, how do the authors explain the relatively low SCS in the Rural railway subset and the modest improvements on attached components?

  2. Could the authors compare PTv3 and HASP under matched sampling and inference protocols? How much of the reported accuracy and efficiency improvement remains when both methods use the same scene-wide sampling configuration?

  3. What cell sizes, minimum track lengths, and distance-decay scales are used for each dataset? How are these parameters adapted to normalized vascular coordinates? Are they determined by a unified scaling rule or tuned separately, and how sensitive are the results to these settings? Cross-scene ablations or comparisons with conventional geometry-aware graph propagation would help clarify the method's generality.

  4. Could the authors provide further validation of SCS and CAS, such as underlying correctness ratios or sensitivity to reference-construction thresholds? Given the relatively small mIoU gains on railway scenes, could additional statistical evidence be provided to establish the robustness of these improvements?

4 comments
AuthorsOpenReview ↗

Response to Reviewer oXVd

Thank you for the constructive comments. We clarify our contributions, module interactions, evaluation protocols, and structural metrics.

  1. Methodological contribution. Earlier methods use structural aggregation and contrastive supervision, but HASP combines them differently. CurveNet learns walks on kNN graphs and CurveCloudNet uses sensor scan polylines, while HASP reconstructs axial chains or surface supports from geometry, separating left, right, and whole-chain context along chains. Unlike Superpoint Graph's region-level propagation, IR combines individual query features, local context, and their difference for pointwise correction. Unlike CBL's boundary-feature contrast, CL supervises the fused query–context features that generate corrections.

  2. Module interactions and CL-only. LR's shared anchor context can recover missed regions but also disturb correct predictions and interrupt continuity. Figure 13 shows correct coverage increasing from 12 to 16 m while the longest correct segment decreases from 12 to 8 m, explaining why higher mIoU can coexist with lower SCS.

IR's geometric query selection can include already-correct points. Ambiguous features may cause incorrect updates because gating controls magnitude rather than class decisions. In Figure 14, four ground-wire points are misclassified as conductors, reducing correct pairs from 30/84 to 23/84. Both endpoints must be correct, so CAS can decrease despite semantic gains.

CL-only reaches 59.82% mIoU and SCS/CAS of 1.79/0.96 in Table 6, with class results in Table 12. With IR, CL supervises fused correction features to resolve ambiguity, recovering three V-insulator tip points in Figure 15. Table 6 shows CAS increasing from 0.95 to 1.08 when CL is added to IR and SCS from 1.84 to 1.92 when LR is added to IR+CL, supporting their complementary roles.

  1. Computational cost. Standard PTv3 uses repeated crop-based inference and voting, while our optimized pipeline processes scene-wide inputs without either. Table 20 reports Full's 61.59% mIoU and estimated 769.72 s/scene, with PTv3's measured 653.22 s/scene as the cost reference. The optimized pipeline achieves 62.27% mIoU in Table 3 and 2.673 s/scene in Figure 6. LR and IR add computation, while the optimized pipeline reduces redundant inference. This overhead represents a reasonable trade-off for improved accuracy and continuity.

  2. Rural continuity and small attachments. Table 17 shows lower rail continuity for HASP than PTv3 at 0.66500 versus 0.67988, but higher wire continuity at 0.98554 versus 0.95337. Figure 17 shows rail-to-track-bed errors leaving a 14 m longest correct segment. Higher rail IoU therefore need not improve continuity.

Small attachments can be geometrically ambiguous. For example, V-insulator arms resemble suspension-insulator segments, so IR may disturb correct LR predictions. Table 6 reports V-insulator IoU of 85.55%, 73.69%, and 81.34% for LR, Joint, and Full. Figure 16 shows three arm endpoints recovered with CL. Despite partial recovery relative to LR, Full improves several attachment classes and raises average attachment IoU by 1.12 points over Base in Tables 14 and 15, supporting the combined modules' effectiveness.

  1. Geometric settings and graph comparison. Table 11 specifies 4 m anchor cells, at least four anchors and 16 m chain length, 64 m directional decay, and 1 m local cells for transmission and railway. IntrA uses centered, unit-radius coordinates, 0.10 local cells, and a 0.20 surface radius. To assess sensitivity, we jointly scale the local cells and surface radius by 0.8, 1.0, and 1.2 on IntrA F0. Table 19 reports 92.59%, 94.67%, and 94.38% mIoU. Defaults perform best, with greater sensitivity to smaller supports.

With the same backbone and inputs, the graph baseline uses three aggregation hops over up to 16 neighbors within 12 m, then local aggregation and gated updates. Table 13 reports Joint/graph mIoU of 61.17%/61.05% and mAcc of 72.14%/68.73%. LR separates directional context, while IR uses feature similarity and geometry for pointwise correction. This combines structural evidence with local discrimination, consistent with higher mean class recall.

  1. Metric validation and railway statistics. To assess SCS/CAS robustness, we fix predictions on 100 scenes and vary correctness thresholds, reference radius, and path step. Table 10 reports Base/Full mC of 0.83992/0.85358 and mX of 0.63119/0.64355. Full remains higher at thresholds of 0.4 and 0.6 and under ±20% radius and step changes, supporting comparison stability across tested settings.

Tables 4 and 16 and Appendix D.3 report railway gains. HASP improves mIoU by 0.44, 0.34, and 0.50 points over PTv3 on Urban, Rural, and Plateau. Table 16 also shows higher mAcc and overhead-line IoU across all subsets, supporting the benefits of structural context.

Reviewer_oXVdOpenReview ↗

Reviewer Response: Unresolved Evidence Gaps on Ablation, OA Protocol, and Methodological Novelty

Thank you for the authors’ response. I appreciate the clarification on the structural inductive bias for HAR scenes, the unified benchmark, and the structural metrics.However, the following issues remain:

In Table 6, the CAS of Full is lower than that of IR+CL, and this has still not been directly explained. The authors explain the improvement in SCS brought by LR, but do not explain why adding LR to IR+CL reduces CAS from 1.08 to 1.03. It is insufficient to explain the coexistence of SCS and mIoU solely by “increased coverage but a shorter longest correct segment”; a direct account of the CAS drop is needed. In addition, the discrepancy in Table 3 between HASP(Base) OA of 95.50% and PTv3 OA of 78.06% has not been addressed. Please clarify whether this discrepancy is caused by the sampling protocol, cropping/full-scene evaluation, class imbalance, or the OA denominator setting. If the protocols differ, the OA difference should not be attributed directly to the method itself. I acknowledge the new benchmark and metrics constructed by the authors, and I support their contribution to model evaluation in this field. However, the advantages of HASP itself across multiple benchmarks are not pronounced, and the explanation of its negative effects remains insufficient. Mechanistically, through the authors’ response and further reading of the paper, I now better understand the novel combination of structural aggregation and contrastive supervision, but I still do not consider this to constitute the main methodological novelty.

I am not inclined to raise my score at this time.

AuthorsOpenReview ↗

Clarifications on CAS regression, OA differences, and methodological contribution

Thank you for the follow-up comments. We have added class-level and visual analyses in Appendices D.2 and D.3.1 to explain the CAS decrease and the OA difference directly.

  1. CAS decrease from IR+CL to Full. To investigate this decrease, we conducted additional class-level evaluations and visual comparisons using the same reference pairs. Table 14 identifies ground wire and optical cable as the main contributors. Their correct-pair ratios decrease from 0.35666 to 0.31788 and from 0.82946 to 0.80212, accounting for 98.94% of the net mX decrease from 0.66026 to 0.64355. This corresponds to the CAS decrease from 1.08 to 1.03 in Table 6.

A possible explanation is that LR’s shared anchor-level context changes the features used by IR in locally ambiguous regions, where ground wires and optical cables can resemble nearby conductors. Figure 16 shows six ground-wire and four optical-cable endpoints correctly classified by IR+CL but misclassified as conductors by Full. Correct pairs decrease from 8/8 to 0/8 and from 8/8 to 1/8, respectively, because one incorrect endpoint can invalidate multiple pairs. These results locate the CAS decrease primarily in local confusion between nearby line classes, while Table 6 shows that adding LR improves overall mIoU from 61.22% to 62.27% and SCS from 1.84 to 1.92.

  1. OA difference between PTv3 and Base. The original configurations use different spatial sampling. PTv3 uses spherical training crops capped at 120,000 points, while Base uses scene-wide samples capped at two million points. Appendix D.3.1 reports a comparison of saved predictions on the same points across 100 scenes, yielding 78.20% OA for PTv3 and 90.39% for Base. Table 17 shows ground recall increasing from 36.82% to 88.11%, vegetation recall from 85.44% to 91.51%, and tower precision from 28.16% to 88.29%. One possible explanation is that scene-wide sampling preserves terrain and structural context that small crops may lack. Ground and vegetation dominate the point population, so their improved recognition strongly affects OA. Figure 18 illustrates fewer ground-to-vegetation and vegetation-to-tower errors. Across these scenes, vegetation-to-tower errors decrease from 66.11 million to 3.28 million points.

  2. Methodological contribution. Although previous studies have explored structural aggregation and contrastive supervision, HASP adapts their construction and interaction to HAR scenes, where long, closely spaced structures coexist with small attachments. As explained in Appendix D.1, CurveNet learns guided walks and CurveCloudNet uses sensor scan polylines, while HASP constructs axial chains or surface supports from input geometry. Along axial chains, LR separately aggregates left, right, and whole-chain context, providing distant structural evidence when local observations are sparse or ambiguous. Superpoint Graph propagates region-level descriptors, whereas IR retains individual query features and combines them with the LR-updated local context and their difference. This allows nearby points to receive different corrections even when they share structural context, helping distinguish attachments from surrounding wires or supports. Unlike CBL’s contrastive supervision of boundary features, our CL directly supervises the fused query-context representation used to generate corrections, encouraging class separation where local aggregation may mix evidence from neighboring classes. Thus, the distinction lies in how geometric supports, pointwise correction, and contrastive supervision work together to address long-range continuity and local semantic ambiguity in HAR scenes.

Reviewer_oXVdOpenReview ↗

Reply to the Authors: Further Explanation of Performance Degradation and Score Adjustment

Thank you to the authors for providing further explanation regarding the performance degradation caused by certain mechanisms. The additional evidence also makes the results in the paper more convincing. I am raising my score from 2 to 3.

Reviewer_42k9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors proposed HASP-Net, a manifold-guided point cloud segmentation framework for high-aspect-ratio structures. Built upon PTv3, it first introduces a Long-range Track Module to preserve structural continuity and a Geometry-guided Interface Module with contrastive learning to improve local boundary discrimination. Then, the authors also introduced HAR3D-Seg, a benchmark covering transmission, railway, and vascular scenes, along with two structural metrics, SCS and CAS. Experiments demonstrate improved segmentation performance and promising inference efficiency compared with existing methods.

Critical strengths
  • The paper addresses the important challenges of long-range semantic inconsistency and boundary confusion in high-aspect-ratio point clouds.
  • Experimental results demonstrate improvements over established baselines, while the reported inference efficiency suggests potential for practical applications.
Critical weaknesses
  • The proposed HASP-Net mainly combines a PTv3 backbone with geometry-based track construction, bidirectional feature aggregation, local attention refinement, and contrastive learning. Most components are based on established techniques, and the paper does not sufficiently demonstrate a fundamentally new learning mechanism.
  • As shown in Tables 4 and 5, although HASP-Net achieves improvements across the evaluated datasets, the gains over PTv3 are relatively small in several settings. Specifically, the mIoU improvements on the three railway subsets are only ~0.4% percentage points, while the vascular dataset shows a ~1% improvement. These modest gains raise questions about whether the additional architectural complexity is justified.
  • The Long-range Track Module relies on geometry-consistent support chains constructed using predefined spatial partitioning and geometric thresholds. However, the paper does not sufficiently investigate its sensitivity to point density, noise, or missing structures. Incorrect track connections may also propagate misleading features between nearby structures belonging to different semantic classes.
  • As shown in Table 6, improvements in overall mIoU do not consistently translate into better segmentation of small attached components. Specifically, the Long-range-only variant achieves 85.55% IoU for the V-ins. class, whereas the complete model (Joint+CL) achieves only 81.34%, despite its higher overall mIoU. This raises concerns about whether the proposed modules consistently benefit fine-grained structural segmentation.
Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 42k9

Thank you for the constructive comments. We have expanded the revised manuscript to clarify the methodological contribution, accuracy–cost trade-off, input robustness, and performance on small attachments.

  1. Methodological contribution. Although previous studies have explored structural aggregation and contrastive supervision, HASP differs in how it constructs supports and combines aggregation, pointwise refinement, and supervision. For example, CurveNet learns guided walks on kNN graphs and CurveCloudNet aggregates along sensor scan polylines, while HASP reconstructs axial chains or surface supports from input geometry and explicitly separates left, right, and whole-chain context along axial chains. Besides, Superpoint Graph propagates region-level descriptors, whereas IR retains individual query features and combines them with local context and their difference for pointwise correction. CBL contrasts features at label-mined boundaries, while our CL directly supervises the fused query–context representation that generates corrections, linking class discrimination to local refinement. In conclusion, although similar techniques have been used before, HASP differs in how structural supports are constructed, pointwise corrections are generated, and contrastive supervision is applied.

    1. Accuracy and computational cost. Table 6 shows that adding LR, IR, and CL to Base increases mIoU from 58.49% to 62.27%, SCS from 1.83 to 1.92, and CAS from 1.00 to 1.03. LR and IR introduce additional computation for context aggregation and pointwise correction. However, Figure 6 shows that our optimized scene-wide pipeline requires 2.673 s per scene, compared with 22.467–653.224 s for the crop-based baselines. The additional computation represents a reasonable trade-off for improved segmentation accuracy and structural continuity.
  2. Density, noise, and missing structures. Coordinate noise has a larger impact than density reduction or local segment deletion. Table 21 shows that noise with standard deviations of 0.02 m and 0.05 m reduces mIoU from 53.55% to 52.57% and 46.94%, respectively. Reducing density to 50% and 25% causes smaller drops of 0.01 and 0.64 percentage points, while deleting 1 m or 2 m segments has little effect on classification of the remaining points. These results suggest that useful structural support remains after thinning or local deletion, while coordinate noise has a greater impact because it can distort estimated directions and neighborhoods.

  3. Incorrect context propagation. Table 22 replaces retained anchor features with features from nearby other chains which introduces 36 errors and reduces mIoU by 0.03 percentage points. Disabling LR removes the effect, while learned gates and gates fixed to 1 give the same loss. This intervention shows that incorrect context can affect otherwise correct predictions through LR and Gating controls update magnitude but does not guarantee that the propagated context supports the correct class. Overall, the impact is small under the tested intervention, but misleading context can still propagate through LR despite gating.

  4. Small attachments. V-insulator arms can locally resemble suspension-insulator segments, making IR corrections ambiguous. Table 6 shows that adding IR to LR reduces V-insulator IoU from 85.55% to 73.69%. CL improves class separation in the correction features, raising IoU to 81.34%. Figure 16 illustrates this recovery through three arm endpoints misclassified by Joint and corrected by Full. Although the recovery is incomplete relative to LR, it supports CL’s role in reducing local confusion during refinement. Overall, compared with Base, Full improves IoU for several small attachment classes and increases average attachment IoU by 1.12 percentage points, as shown in Tables 14 and 15. These results support the effectiveness of the combined modules for small-attachment segmentation.

AuthorsOpenReview ↗

Further clarification of the HAR-specific design and local prediction errors

Following our previous response, we have clarified how the design addresses HAR scenes and added a class-level analysis of the remaining local prediction errors. The additional material has changed the numbering in the revised manuscript. Tables 14, 15, 21, and 22 cited in our previous response are now Tables 15, 16, 23, and 24, respectively. Figure 16 in that response is now Figure 17. All references below follow the latest revision.

  1. Methodological contribution in HAR scenes. Although previous studies have explored structural aggregation and contrastive supervision, HASP adapts their construction and interaction to long, closely spaced structures and small attachments. As explained in Appendix D.1, CurveNet learns guided walks and CurveCloudNet uses sensor scan polylines, while HASP constructs axial chains or surface supports from input geometry. LR separately aggregates left, right, and whole-chain context along axial chains, providing distant evidence when local observations are sparse or ambiguous. Superpoint Graph propagates region-level descriptors, whereas IR retains individual query features and combines them with local context and their difference for pointwise correction. This allows nearby points sharing structural context to receive different corrections. Unlike CBL's supervision of boundary features, our CL supervises the fused query-context representation used to generate corrections, encouraging class separation where local aggregation may mix neighboring classes. The distinction lies in how these choices jointly address long-range continuity and local semantic ambiguity in HAR scenes.

  2. Additional evidence on local confusion and context propagation. Appendix D.2 examines prediction changes from IR+CL to Full using fixed evaluation pairs. Table 14 shows that ground wire and optical cable account for 98.94% of the net mX decrease, with correct-pair ratios changing from 0.35666 to 0.31788 and from 0.82946 to 0.80212. Figure 16 shows six ground-wire and four optical-cable endpoints changing from correct predictions to the conductor class. Correct pairs consequently decrease from 8/8 to 0/8 and from 8/8 to 1/8. These observations are consistent with shared context affecting locally ambiguous predictions. This analysis complements the context-injection experiment in Table 24 by identifying local errors in the original model predictions.

  3. Overall and small-attachment benefits. Under shared sampling and evaluation settings, Table 6 reports a Base-to-Full mIoU increase from 58.49% to 62.27%. For small attachments, Table 16 shows improvements over Base for strain insulators, V-insulators, and suspension insulators. Table 15 reports an increase in average attachment IoU from 87.45% to 88.57%. Figure 17 illustrates three V-insulator arm endpoints corrected by Full after being misclassified by Joint, consistent with the V-insulator IoU recovery from 73.69% to 81.34% in Table 6. Full remains below LR's 85.55% on this class, but these results support improved average attachment segmentation and CL's role in reducing local confusion during refinement.

Reviewer_E1XROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HASP-Net for semantic segmentation of elongated point-cloud structures and their attached components. Based on PTv3, it introduces geometric track-based context aggregation, local interface refinement and contrastive learning. The paper also presents HAR3D-Seg dataset, which combines newly collected transmission-line data with existing railway and vascular data, and reports improved mIoU across the three domains.

Critical strengths
  • The newly proposed transmission-line dataset is large and includes fine-grained annotations for small attached components, providing a useful benchmark for this problem. -The proposed SCS and CAS take long-range semantic continuity and local class separation into consideration, which complements standard point-wise segmentation metrics.
  • The proposed method improves mIoU across transmission, railway and vascular scenes, showing its effectiveness for different types of elongated structures.
Critical weaknesses
  • The method is tailored to elongated structures, and its broader applicability is unclear. The geometric settings used across the three domains are not fully explained.
  • The structural benefits are not clear from the ablations. LR alone does not improve SCS, IR alone lowers CAS, and Base+CL is missing.
  • The source of the speedup needs clarification. HASP and the baselines use different inference pipelines, and the runtime of Base is missing.
Questions for authors
  • How does track-based aggregation compare with graph-based aggregation using the same backbone and inputs?
  • Why do LR and IR alone not improve their intended structural metrics? What are the results of Base+CL?
  • What are the runtimes of base setting and full under the same scene-wide inference setting?
  • What geometric settings are used for railway and vascular data, and how are they selected?
2 comments
AuthorsOpenReview ↗

Response to Reviewer E1XR

Thank you for the constructive comments. We have expanded the revised manuscript to clarify the geometric settings, module interactions, and computational cost.

  1. Generalization. Table 18 extends the evaluation to DALES and S3DIS, covering general outdoor and indoor scenes. Under same settings within each dataset, Full maintains baseline accuracy with small gains. These results support its applicability beyond the targeted high-aspect-ratio scenes, although the additional benefit in general scenes is limited.

2.Geometric settings. Table 11 specifies the parameters across domains. Transmission and railway share 4 m anchor cells, a minimum chain length of 16 m, and a 64 m directional decay scale. IntrA uses centered, unit-radius coordinates with a local-cell size of 0.10 and a surface-support radius of 0.20. Table 19 further examines sensitivity to changes in the local-cell size and surface-support radius.

  1. Module interactions and CL-only. LR provides long-range semantic evidence, but points assigned to the same anchor share aggregated context which may recover missed regions while disturbing locally correct predictions. Even a few new errors can interrupt continuity. In Figure 13, LR increases correct coverage from 12 to 16 m but shortens the longest correct segment from 12 to 8 m, explaining why higher mIoU can coexist with lower SCS.

IR selects queries geometrically, so it may also update already-correct points. Ambiguous local features can produce incorrect updates because the gate controls the size rather than the class decision. In Figure 14, four ground-wire points become misclassified as conductors, reducing correct pairs from 30/84 to 23/84. Since both endpoints must be correct, these errors can lower CAS despite overall semantic gains.

Table 6 reports CL-only’s 59.82% mIoU and SCS/CAS of 1.79/0.96, with class-level results in Table 12. With IR, CL supervises the fused correction features, helping resolve ambiguity, as illustrated by three recovered V-insulator tip points in Figure 15. Table 6 further shows that adding CL to IR raises CAS from 0.95 to 1.08, while adding LR to IR+CL raises SCS from 1.84 to 1.92, supporting their complementary roles.

  1. Graph comparison. Using the same backbone and inputs, the graph baseline performs three aggregation hops over up to 16 neighbors within 12 m, followed by local aggregation and gated updates. Table 13 shows similar mIoU for Joint and the graph baseline at 61.17% and 61.05%, while Joint achieves higher mAcc at 72.14% versus 68.73%. Unlike neighborhood-based graph aggregation, LR explicitly separates ordered left, right, and whole-chain context. IR then uses feature similarity and geometry for pointwise correction. This combination retains directional structural evidence while resolving local class ambiguities, consistent with the improved mean class recall.

  2. Computational cost. We distinguish two inference pipelines: standard PTv3 with repeated crop-based inference and voting, and our optimized scene-wide pipeline without repeated cropping or voting. Table 20 reports Full’s 61.59% mIoU and an estimated runtime of 769.72 s/scene, using PTv3’s measured 653.22 s/scene as the cost reference. With our optimized pipeline, Full achieves 62.27% mIoU in Table 3 and requires 2.673 s/scene in Figure 6. LR and IR introduce additional computation for structural aggregation and local correction, while the optimized pipeline reduces redundant inference and voting. Although the modules increase computational cost, this overhead represents a reasonable trade-off for improved segmentation accuracy and structural continuity.

AuthorsOpenReview ↗

Additional evidence on module interactions and evaluation settings

Following our previous response, we have added class-level and visual analyses in Appendices D.2 and D.3.1 to further explain the structural results and the effect of sampling settings. The additional material has changed the numbering in the revised manuscript. Tables 18, 19, and 20 cited in our previous response are now Tables 20, 21, and 22, respectively. All references below follow the latest revision.

  1. Further analysis of module interactions. To investigate why adding LR to IR+CL improves SCS but lowers CAS, we conducted additional class-level evaluations using the same reference pairs. Table 14 identifies ground wire and optical cable as the main contributors. Their correct-pair ratios decrease from 0.35666 to 0.31788 and from 0.82946 to 0.80212, accounting for 98.94% of the net mX decrease from 0.66026 to 0.64355.

A possible explanation is that LR's shared context changes the features used by IR where nearby line classes remain locally ambiguous. Figure 16 shows six ground-wire and four optical-cable endpoints correctly classified by IR+CL but predicted as conductors by Full. Correct pairs decrease from 8/8 to 0/8 and from 8/8 to 1/8, respectively, because one incorrect endpoint can invalidate multiple pairs. This localizes the CAS decrease from 1.08 to 1.03 to specific class confusions, while Table 6 shows higher mIoU from 61.22% to 62.27% and SCS from 1.84 to 1.92. The combined modules therefore improve overall accuracy and continuity without improving every local interface equally.

  1. Sampling settings and OA differences. PTv3 uses spherical training crops capped at 120,000 points, while Base uses scene-wide samples capped at two million points. Appendix D.3.1 compares saved predictions on the same original points across 100 scenes, yielding 78.20% and 90.39% OA, respectively. Table 17 reports ground recall increasing from 36.82% to 88.11%, vegetation recall from 85.44% to 91.51%, and tower precision from 28.16% to 88.29%. Scene-wide sampling may preserve terrain and structural cues that small crops lack. Since ground and vegetation dominate the point population, their improved recognition strongly influences OA. Figure 18 illustrates fewer ground-to-vegetation and vegetation-to-tower errors, with the latter decreasing from 66.11 million to 3.28 million points across these scenes.

  2. Module benefits under shared settings. Under shared sampling and evaluation settings, Table 6 shows that Full improves Base's mIoU from 58.49% to 62.27%, SCS from 1.83 to 1.92, and CAS from 1.00 to 1.03. Table 22 additionally reports 61.59% mIoU for Full compared with 58.98% for the PTv3 baseline. The Base-to-Full ablation supports the benefits of the combined modules within a consistent evaluation setting.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.