acceptodds
Under review as a conference paper at ICLR 2027

3D Generalized Category Discovery

Abstract

Generalized category discovery (GCD) aims to recognize known categories and discover novel ones in unlabeled data, using supervision from known classes. 3D generalized category discovery (3DGCD) extends this setting to point clouds. Our analysis of pretrained 3D features reveals two problems. Feature neighborhoods contain objects from different categories and change across augmented views. Even when neighborhoods are relatively pure, known objects can dominate clustering and leave novel categories without distinct groups. We propose Geo3D to address both problems by coupling geometric relations between objects with category-level guidance. Geometric compatibility and agreement across views help identify reliable neighbors. Geo3D combines their relational evidence with a memory of category structure to construct discovery targets. Objects confidently assigned to known categories are anchored to their predicted classes, reducing their competition for novel groups during training. The resulting supervision supports coherent grouping while allowing novel categories to form distinct clusters. Extensive experiments on synthetic shapes and real-world scans demonstrate the effectiveness of the proposed method.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generalized Category Discovery, 3D Point Clouds, Open-World Recognition

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wqxnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Geo3D for point-cloud generalized category discovery, combining geometric neighborhood refinement with memory-guided category assignment and known-class anchoring.

Critical strengths

First, Geo3D achieves the highest overall and novel-class accuracy across three benchmarks, improving novel-class accuracy over the strongest competitors by 5.69–7.00%. Second, results with two backbones and multiple random seeds strengthen the empirical evidence beyond a single configuration.

Critical weaknesses

(1) In Table 11, adding MGCA reduces overall accuracy on all three datasets before KnownEx and EMA are introduced. The full ScanObjectNN configuration also adds dataset-specific gates. Although leave-one-out ablations are provided, a precise baseline definition and gate-controlled comparisons are needed to clarify these interactions. (2) Table 2 clusters the mixed known–novel pool into only the number of novel categories. Consequently, low novel coverage may partly reflect the restricted cluster budget. Repeating the diagnostic with all KK clusters would better establish its relevance to the evaluation setting.

Questions for authors

Please consider my two main concerns: (1) Does Geo3D also outperform the strongest baselines in novel-class macro accuracy? On ShapeNetCore, its 69.98% micro accuracy contrasts with 44.73% macro accuracy, leaving improvements in rare-category discovery unclear. (2) What is the end-to-end training cost, including memory refresh, frozen-reference extraction, and prototype updates, which are excluded from the reported forward FLOPs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer wqxn

Thank you. Dataset order: MN=ModelNet40, Scan=ScanObjectNN, Shape=ShapeNetCore.


◆ Q1: Does novel macro accuracy support gains over baselines and rare classes on Shape?

A1: Supplement: Ward with K=40/15/55 and one global Hungarian mapping shared by All/Old/New/macro. Main-table predictions remain native.

Table R1a: Novel-class macro accuracy (%), Ward.

MethodMNScanShape
DGCNN43.2926.9439.95
Point-BERT42.6020.2337.35
Uni3D49.6946.6240.50
LPS53.1056.3243.14
SimGCD53.6253.5838.88
CMS52.9544.6842.40
SelEx57.6251.0141.89
ProtoGCD53.6855.9944.46
CryoGCD27.3733.4527.95
Baseline48.7752.8242.70
Geo3D58.0763.4644.73

Geo3D leads listed methods by 0.45/7.14/0.27 points over SelEx/LPS/ProtoGCD, respectively. Gains over Baseline are 9.30/10.64/2.03 points; Shape also gains 4.23 over frozen Uni3D. Shape's 69.98% micro weights samples equally; 44.73% macro averages per-class accuracy, equally weighting rare classes and supporting better average novel-class recognition. Appendix L (Table 23) is updated in purple.


◆ Q2: What are training time/peak GPU memory, including refresh, reference inference and prototype updates?

A2: All methods use Uni3D, 1,024 points; timing uses one RTX 4090, batch size 16 and native training/evaluation procedures. Geo3D includes memory refresh, reference-feature extraction, geometry/target construction, parameter updates, evaluation and checkpoint saving.

Table R2a: Trainable parameters (M) / single-object forward FLOPs (G).

MethodMNScanShape
BIA8.45/88.488.42/88.488.46/88.48
ProtoGCD13.95/87.6913.93/87.6913.97/87.69
SimGCD13.95/87.6913.93/87.6913.97/87.69
Geo3D9.72/87.6810.25/87.699.72/87.68

Table R2b: Measured minutes per epoch.

MethodMNScanShape
BIA3.113.7211.49
ProtoGCD4.695.3816.53
SimGCD4.314.8615.57
Geo3D3.393.9112.45

Geo3D: 100-epoch times 5.65/6.52/20.75 hours; per-epoch peak allocated GPU tensor memory is 3.84/3.88/4.46 GiB. It improves novel discovery with 26.4%–30.4% fewer trainable parameters, comparable forward computation and lower per-epoch time than ProtoGCD/SimGCD on all datasets. Appendix B.4, Tables 8(a)/(b), adds costs/conditions in dark green, shared by both reviewers.


◆ Q3: Why the Table 11 drop at +MGCA? What are Baseline and Scan-specific settings?

A3: Baseline: Uni3D-B, contrastive objectives, prototype supervision, global/local representations, neighborhood consensus; 3DGR/MGCA/KnownEx/EMA and Scan-specific settings are off. Table 11 adds modules successively; space splitting/gradient projection start at +MGCA (§3.5). Scan Full adds robust geometry processing/dynamic reliability gating, separate from all datasets’ robust pair-scoring descriptor.

Table 3 removes MGCA within Full; both Scan configurations retain robust geometry and gating.

Table R3a: MGCA removal, All/Old/New (%).

Config.MNScanShape
w/o MGCA77.84/89.93/66.3569.50/73.69/63.1078.83/86.72/62.19
Full78.40/90.77/66.6770.68/75.53/63.2880.61/85.65/69.98

Full gains 0.56/1.18/1.78 All and 0.32/0.18/7.79 New points.

Table R3b: Novel objects with a known-slot historical argmax in training memory (%).

Dataset+MGCA+KnownExFull
MN30.13%25.86%21.72%
Scan45.51%34.70%26.10%
Shape29.58%24.34%24.68%

+MGCA→Full reduces this confusion by 8.41/19.41/4.90 points, complementing Table 3.

Table R3c: Scan configurations, accuracy (%).

GeometryGatingAllOldNew
OffOff66.9775.7653.55
OnOff68.0171.9761.96
OnOn (Full)70.6875.5363.28

Geometry alone gains 1.04 All/8.41 New points over both off; Full gains another 2.67 All/1.32 New points. These configurations supplement the ablation. Main text/Appendix J (Tables 20–21) updated in purple.


◆ Q4: Does restricted K explain Table 2's low coverage? What changes with total K?

A4: We use K=K_old+K_new, fixing frozen features, diagnostic samples, natural-class-frequency weights and other weighted K-Means settings. Coverage: fraction of novel classes becoming the largest individual class in ≥1 novel-dominated cluster.

Table R4a: Novel coverage (%) by cluster budget.

DatasetClustersOriginalTotal K
MN20→4060.00%85.00%
Scan7→1571.43%71.43%
Shape27→5525.93%51.85%

Total K covers 17/20, 5/7 and 14/27 novel classes: coverage rises on MN/Shape and stays unchanged on Scan. Known-class weighted fractions within novel-dominated clusters are 15.64%/17.54%/15.13%, complementing coverage. This comparison shows the cluster-budget effect. Main text/Appendix K (Table 22) updated in purple.

Reviewer_wqxnOpenReview ↗

Review

Thank authors for your feedback. Some of my concerns are addressed, so I would like to raise my score.

AuthorsOpenReview ↗

Response to Reviewer wqxn

We sincerely thank the reviewer for the thoughtful reassessment and for raising the score. We greatly appreciate your time, constructive feedback, and recognition of our efforts to address the concerns.

Reviewer_8itxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Geo3D aims at the unreliable neighborhood and known-class dominance problems in 3D Generalized Category Discovery, combining 3D Geometry Reliability (3DGR) and Memory-Guided Category Assignment (MGCA), using geometric relationships and category memory to guide feature learning. The paper reports better All and New accuracy on three data sets, but some mechanism definitions and experimental attributions still need to be clarified.

Critical strengths

1.The method revolves around two issues: neighbor reliability and category assignment, and the design motivation is clear.

2.The experiments are rich, including component ablation, graph intervention, different backbone and multi-seed results, and the performance trade-off between known and novel categories is analyzed.

Critical weaknesses

1.The definition of key mechanisms is incomplete. Section 3.4 and Appendix A do not clearly provide the sample retention rules of KnownEx, the threshold calibration method and margin of known acceptance, as well as the complete calculation of edge factor a_{ij} and object reliability r_i. These quantities directly affect neighbor weighting and category assignment. Without complete definitions, it is difficult to independently reproduce the method from the current description.

2.The configuration boundaries between Baseline and component ablation are unclear. Table 11 does not fully explain which auxiliary losses, prototype supervision and gradient projection are retained by Baseline. Appendix C.3 also points out that the Full configuration of ScanObjectNN contains dataset-specific reliability gates, so it is difficult to clearly attribute some of the gains to 3DGR, MGCA or additional gating mechanisms.

3.The paper lacks controlled comparisons that isolate the contributions of representation learning and the clustering algorithm. Section 4.1 explains that Geo3D uses Ward clustering and frozen representations use K-Means. Therefore, the improvement relative to frozen baselines includes changes in both representation learning and clustering algorithms, and the current results cannot quantify the contributions of both separately.

Questions for authors

1.Table 2 clusters mixed data containing known and novel categories into clusters with the number of novel categories, and the main experiment uses the total number of categories K. Are known-class dominance and low Novel Coverage still evident when using K clusters? Can MGCA improve the same diagnostic metrics in this setting?

2.What exactly are replaced by Constant and Shuffled in Figure 3: geometric descriptors, pair scores, geometric correction, or final graph weights? Please indicate whether feature similarity and other mechanisms remain unchanged in these comparisons, so as to determine whether the experiment independently tests the contribution of geometric information.

3.Could you evaluate both frozen features and Geo3D features using Ward clustering, then repeat the comparison using K-Means? This would help separate the gains from feature learning and clustering. Please also clarify how the main GCD baselines in Table 1 produce their final predictions.

I will consider raising the rating based on the author's clarification and additional evidence.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 8itx

Thank you. Results (%), dataset order: MN/Scan/Shape=ModelNet40/ScanObjectNN/ShapeNetCore.


◆ Q1: Total-K diagnostics?

A1: Table 2 weights/protocol; budgets 20→40/7→15/27→55. Coverage: fraction of novel classes largest in ≥1 novel-dominated cluster. F=frozen Uni3D; T=trained with MGCA.

Table R1a: Coverage.

DataF: originalF: KT: K
MN60.0085.0085.00
Scan71.4371.4385.71
Shape25.9351.8566.67

Table R1b: K, F→T.

DataOld-dominatedOld mass in novel-dominated↓Purity↑
MN35.00→30.0015.64→10.2475.54→79.27
Scan40.00→33.3317.54→15.1954.66→71.30
Shape52.73→38.1815.13→7.8979.75→88.50

Larger K raises F’s MN/Shape coverage. T: MN coverage unchanged, Scan/Shape +1/+4 classes; old mass↓, purity↑. Removing MGCA from Full lowers All 0.56/1.18/1.78, New 0.32/0.18/7.79 (Table 3). Updated Appendix K/Table 22; P/Tables 27–28 (red).


◆ Q2: Constant/Shuffled?

A2: Cross-view support: Constant=row mean; Shuffled=one-position cyclic shift (distribution preserved). Geometry/pair scores/supervision/similarity rules stay; features/neighbors/weights may change. This tests cross-view support.

Table R2a: Figure 3.

SupportAllOldNew
Correct70.6875.5363.28
Constant64.7173.6951.01
Shuffled64.8971.2855.13

Correct: New +12.27/+8.15 points.

Table R2b: New.

RemovedMNScanShape
3DGR66.2750.2264.59
Pair loss64.7757.1465.77
None66.6763.2869.98

Drops 0.40/13.06/5.39 and 1.90/6.14/4.21 support module/pair-supervision contributions. Updated Figure 3/O/Tables 25–26.


◆ Q3: Clustering/predictions?

A3: Total K; K-Means seed 42/n_init=10.

Table R3a: All/Old/New.

ClustererFeaturesMNScanShape
K-MeansF59.76/69.05/50.9550.24/47.85/53.9043.76/40.16/51.35
K-MeansBaseline72.97/89.43/57.3563.32/75.42/44.8777.23/84.15/62.65
K-MeansGeo3D75.93/91.60/61.0666.13/74.04/54.0880.17/84.67/70.68
WardF65.52/75.54/56.0050.28/51.35/48.6446.73/39.66/61.65
WardBaseline74.43/90.93/58.7766.52/72.66/57.1478.30/84.44/65.34
WardGeo3D78.40/90.77/66.6770.68/75.53/63.2880.61/85.65/69.98

Geo3D beats both controls in All/New. All gains vs F: K-Means 16.17/15.89/36.41; Ward 12.88/20.40/33.88. Fixed Geo3D features: Ward gains All 2.47/4.55/0.44; New 5.61/9.20, but Shape 70.68→69.98. Both matter; Ward’s effect varies.

Native: frozen/GCD/SelEx=K-Means; SimGCD/ProtoGCD/BIA=head argmax; CMS/Geo3D=Ward(K). One global Hungarian mapping serves All/Old/New. Updated H/Table 18/setup.


◆ Q4: Definitions?

A4: Appendix M/§3.4: normalized features/prototypes; NUN_U=unlabeled views, K=Ko+KnK=K_o+K_n, known-class labeled fraction η=.5η=.5; retention ρk=Kn/(Kn+.5Ko)ρ_k=K_n/(K_n+.5K_o)≠ramp ρρ. Keep M lowest max-known cosine scores s: M=min⁡(NU,max⁡(20Kn,⌈ρkNU⌉)),ω=max⁡(.05,σ((s∗−s)/.05)).M=\min(N_U,\max(20K_n,⌈ρ_kN_U⌉)), ω=\max(.05,σ((s_*-s)/.05)). s∗s_*=largest retained score; σ=sigmoid. Acceptance: student/reference known labels agree, both meet separate classwise 10th labeled-similarity percentiles, student known−novel similarity≥−.02. Accepted probabilities: known-only; others: all K (order in M).

CC clips to [0,1]; A(p,q)=C(1−JSD(p,q)/ln⁡2)A(p,q)=C(1-\mathrm{JSD}(p,q)/\ln2); c=normalized confidence (M). ri=max⁡(.05,C(max⁡(10−8,visvirA(πi,πir)A(πi,hi)(.25+.75ci))1/5)).r_i=\max(.05,C(\max(10^{-8},v_i^sv_i^rA(π_i,π_i^r)A(π_i,h_i)(.25+.75c_i))^{1/5})). v=view stability, π=assignment, h=history (momentum .8); reference temp=.1; missing-history agreement 1; labeled r=1. aij=.1+.9[max⁡(10−8,uijdijrirj)]1/3.a_{ij}=.1+.9[\max(10^{-8},u_{ij}d_{ij}\sqrt{r_ir_j})]^{1/3}. u/d=reference/assignment support (M). Normalize aexp⁡(e)a\exp(e) over 3DGR neighbors. Object values: shared across views, fixed per epoch; gradients through e.


◆ Q5: Baseline/stages/Scan extras?

A5: Baseline: Uni3D-B, contrastive/prototype/auxiliary losses (N). S=shared-space supervision/gradient projection; Scan=robust geometry/dynamic gating ≠ shared descriptors.

Table R5a: Table 11 settings.

Config3DGRMGCAKnownExEMASScan
Baseline——————
+3DGR✓—————
+MGCA✓✓——✓—
+KnownEx✓✓✓—✓—
Full✓✓✓✓✓Scan only

S starts at +MGCA (§3.5); Table 11 adds; Table 3 removes modules:

Table R5b: All/Old/New.

RemovedMNScanShape
3DGR77.39/89.10/66.2766.07/76.45/50.2277.97/84.31/64.59
MGCA77.84/89.93/66.3569.50/73.69/63.1078.83/86.72/62.19
None78.40/90.77/66.6770.68/75.53/63.2880.61/85.65/69.98

Both lower All/New: Scan New−13.06 without 3DGR, Shape New−7.79 without MGCA. Full/−MGCA retain Scan extras.

Table R5c: Scan.

Robust geometryDynamic gatingAllOldNew
OffOff66.9775.7653.55
OnOff68.0171.9761.96
OnOn (Full)70.6875.5363.28

Robust-only: All/New +1.04/+8.41; Full adds 2.67/1.32, supporting ablations. Updated N/Table 24, J/Table 21, Table 11.

Thank you for your response. You noted that both “Full” and “w/o MGCA” preserve the robust geometry and dynamic gating mechanisms from ScanObjectNN. My question is whether “w/o 3DGR” also preserves these two mechanisms in Table R5b.

AuthorsOpenReview ↗

Response to Reviewer 8itx

Thank you for your follow-up question.


◆: Thank you for your response. You noted that both “Full” and “w/o MGCA” preserve the robust geometry and dynamic gating mechanisms from ScanObjectNN. My question is whether “w/o 3DGR” also preserves these two mechanisms in Table R5b.

A: The original “w/o 3DGR” configuration on ScanObjectNN in Table R5b did not retain both robust geometry processing and dynamic reliability gating. Our previous statement referred specifically to Full and “w/o MGCA,” which retain both mechanisms. For completeness, we clarify the configuration of “w/o 3DGR” below.

To examine the contribution of 3DGR with both mechanisms retained, we additionally evaluated the following configuration:

Table R5d: ScanObjectNN accuracy (%) with robust geometry processing and dynamic reliability gating retained.

ConfigurationRobust geometryDynamic gatingAllOldNew
FullEnabledEnabled70.6875.5363.28
w/o 3DGREnabledEnabled69.5474.8461.44

Compared with the reported Full results, All, Old, and New accuracy decrease by 1.14, 0.69, and 1.84 percentage points, respectively. These results support a contribution from 3DGR when both additional mechanisms remain enabled, with a larger difference in novel-category accuracy.

The original Table R5b and this additional experiment represent different configuration settings. We retain the original results and report the comparison with both mechanisms enabled separately to make this distinction explicit.

We have clarified the original configuration in the main-text ablation discussion and the caption of Table 3, and added the comparison and analysis to Appendix Q (Table 29). These manuscript updates are highlighted in red.

Reviewer_xHXKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates 3D Generalized Category Discovery (3DGCD), which aims to recognize known categories and discover novel categories from partially labeled 3D point cloud data. The authors identify two main challenges: unreliable feature neighborhoods caused by geometric ambiguity and augmentation-induced changes, and the dominance of known categories during clustering, which can prevent novel categories from forming distinct groups. To address these challenges, the paper proposes Geo3D, consisting of two main components. The 3D Geometry Reliability (3DGR) module refines neighborhood relationships using geometric descriptors and cross-view agreement. The Memory-Guided Category Assignment (MGCA) module combines neighborhood information, feature memory, and category prototypes to construct discovery targets while reducing interference from known categories. The proposed framework is evaluated on ModelNet40, ScanObjectNN, and ShapeNetCore. The reported results show improvements in overall and novel-category accuracy over the compared methods. The paper also provides component and loss ablations, graph-evidence interventions, multi-seed evaluations, and sensitivity analyses.

Critical strengths
  1. Clear problem motivation and meaningful analysis. The paper identifies two relevant challenges in 3D category discovery. I find the distinction between unreliable neighborhood relationships and known-category dominance reasonable. In particular, the observation that relatively high neighborhood purity does not necessarily lead to good novel-category coverage provides useful motivation for considering category assignment alongside representation learning.
  2. Strong experimental results on multiple benchmarks. Geo3D achieves competitive performance on all three evaluated datasets. In Table 1, its novel-category accuracy exceeds the strongest reported competing results by 5.69, 6.66, and 7.00 percentage points on ModelNet40, ScanObjectNN, and ShapeNetCore, respectively. These improvements are meaningful given the importance of novel-category recognition in GCD.
  3. Comprehensive ablation studies and useful mechanism analysis. The experimental analysis is a strength of this paper. Beyond component-wise and incremental ablations, the authors investigate individual loss terms, graph-evidence interventions, multiple random seeds, and hyperparameter sensitivity. I particularly appreciate the comparison between correct, constant, and shuffled graph evidence in Figure 3. Together with the neighborhood diagnostics in Table 2 and the representation analysis in Table 9, these experiments provide useful evidence supporting the role of reliable neighborhood relationships. The ablation results also show that 3DGR and MGCA are particularly beneficial on datasets exhibiting their corresponding challenges.
  4. Availability of the main implementation. The supplementary material includes implementations of the principal components, including geometric neighborhood construction, memory-guided assignment, and the training objectives. This improves the transparency of the proposed framework.
Critical weaknesses
  1. The methodological novelty relative to existing GCD approaches could be better established. Although Geo3D is reasonably motivated, several underlying ideas have been explored in prior GCD research. For example, AMEND and CMS investigate neighborhood-based representation learning, while ProtoGCD and TAN address prototype-based learning and biases between known and novel categories. Geo3D introduces geometry-aware reliability estimation and a memory-guided assignment mechanism tailored to 3D point clouds. However, the paper could more clearly explain how its proposed mechanisms differ from these existing approaches and what additional benefit is provided by their joint design. I do not consider the integration of existing ideas inherently problematic, especially when it addresses a meaningful problem and produces substantial empirical gains. Nevertheless, a more explicit technical comparison with the closest prior work would strengthen the contribution.
  2. The adaptation and evaluation protocols of the competing methods require further clarification. Table 1 includes methods originally developed for different settings, including 3D representation learning, image-based GCD, 3D novel class discovery, and other 3D data modalities. Evaluating adapted methods on a common 3DGCD benchmark is reasonable, but the adaptation procedures are not sufficiently clear. In particular, it would be helpful to know which components of the original methods were retained or modified when adapting them to point clouds, and whether the methods were evaluated under comparable training and clustering protocols. Section 4.1 also states that Geo3D uses fixed-K Ward clustering, whereas frozen representation baselines are evaluated using K-Means. This is not necessarily an unfair comparison, since different methods may have different inference procedures. However, clarification of the final clustering protocols, and ideally a controlled comparison where appropriate, would help distinguish improvements due to learned representations from those influenced by post-processing.
  3. The computational cost analysis does not fully characterize the complete training pipeline. The paper reports favorable trainable parameter counts and model-level computational costs. However, Geo3D also requires memory-bank updates, frozen-reference feature extraction, geometric descriptor computation, and prototype updates. As acknowledged in Appendix B.4, these additional operations are not fully reflected in the reported forward FLOPs. End-to-end training time and peak GPU memory measurements would provide a more complete assessment of the computational trade-offs.
  4. The title could more clearly reflect the specific contribution. The current title, "3D Generalized Category Discovery," is rather broad and primarily describes the research task. Since the paper specifically focuses on geometry-reliable neighborhood construction and memory-guided category assignment, a more descriptive title would better communicate its contribution. This is a minor presentation concern rather than a substantive technical weakness.
Questions for authors

Q1. Could the authors clarify how the methods originally developed for image-based GCD, 3D novel class discovery, and other 3D data modalities were adapted to the proposed benchmark? In particular, were these methods evaluated under comparable backbone, training, and final clustering protocols? Could the authors also clarify the effect of Ward clustering versus K-Means on the reported results? Q2. How does Geo3D differ technically from existing neighborhood-based GCD approaches such as AMEND and CMS, and prototype-based approaches such as ProtoGCD and TAN? In particular, what is the additional contribution of jointly using geometry-aware neighborhood refinement and memory-guided assignment, beyond combining these established ideas? Q3. Could the authors report the end-to-end training time and peak GPU memory usage under a clearly specified hardware configuration, including the additional overhead from memory refresh, frozen-reference inference, and prototype updates?

2 comments
AuthorsOpenReview ↗

Response to Reviewer xHXK

Thank you for your comments. We have revised the title to reflect our method.


◆ Q1: Method adaptation, protocol comparability and Ward/K-Means?

A1: Adaptation/clustering details are added in blue (Appendix H, Table 18).

Same 3D GCD task/splits. Image methods use pretrained Uni3D-B, point-cloud augmentations and object losses. SimGCD keeps supervised classification/self-distillation/contrast; ProtoGCD keeps prototype learning/separation. Novel-discovery methods use mixed known/novel unlabeled pools and all-class outputs. Other-modality encoders are replaced; CryoGCD keeps semantic prototype alignment/two-stage learning with Uni3D. Learning mechanisms remain; frozen baselines evaluate pretrained features.

SimGCD/ProtoGCD/BIA use classifier/prototype-head argmax; GCD/SelEx use native K-Means; CMS/Geo3D use total-K Ward. One global Hungarian mapping evaluates predictions for All/Old/New.

Both clusterers use total K. Supplementary K-Means: seed 42, n_init=10. Entries: All / Old / New (%).

Table R1a: Clustering comparison.

ClusteringFeaturesModelNet40ScanObjectNNShapeNetCore
K-MeansFrozen Uni3D59.76 / 69.05 / 50.9550.24 / 47.85 / 53.9043.76 / 40.16 / 51.35
K-MeansBaseline72.97 / 89.43 / 57.3563.32 / 75.42 / 44.8777.23 / 84.15 / 62.65
K-MeansGeo3D75.93 / 91.60 / 61.0666.13 / 74.04 / 54.0880.17 / 84.67 / 70.68
WardFrozen Uni3D65.52 / 75.54 / 56.0050.28 / 51.35 / 48.6446.73 / 39.66 / 61.65
WardBaseline74.43 / 90.93 / 58.7766.52 / 72.66 / 57.1478.30 / 84.44 / 65.34
WardGeo3D78.40 / 90.77 / 66.6770.68 / 75.53 / 63.2880.61 / 85.65 / 69.98

Geo3D exceeds Baseline in All/New throughout. New gains (dataset order): 3.71/9.20/8.03 points with K-Means and 7.90/6.14/4.64 with Ward.


◆ Q2: Technical distinctions and joint 3DGR/MGCA contribution?

A2: Added in blue: Appendix I, Table 19.

Table R2a: Technical comparison.

MethodMain mechanismGeo3D distinction
AMENDExpanded-neighborhood contrast; class-wise adaptive prototype angular margins.Geometry/cross-view neighbor scoring and selection.
CMSMean-shift aggregation, then shifted-feature contrast.Refined neighbors + object memory/prototypes/reliability form targets.
ProtoGCDJoint prototypes; dual-level adaptive known/novel pseudo-labeling.Refined neighbors + prototype targets; history/reliability weight guidance.
TANKnowledge transfer, prototype calibration, instance/category alignment via known prototypes.Geometry/cross-view neighborhoods feed memory/reliability-guided targets.

3DGR scores/selects neighbors via covariance eigenvalues, radial statistics, multi-scale local descriptors and cross-view support. MGCA keeps these neighbors/scores, using object memory/prototypes/reliability to adjust weights/targets. Reliable objects get stronger neighbor/prototype guidance; uncertain ones keep more current feature direction. 3DGR selects relationship evidence; MGCA controls its category-target influence.

Table 3 removes modules from Full:

Table R2b: Module ablations, All/Old/New (%).

ConfigurationModelNet40ScanObjectNNShapeNetCore
w/o 3DGR77.39 / 89.10 / 66.2766.07 / 76.45 / 50.2277.97 / 84.31 / 64.59
w/o MGCA77.84 / 89.93 / 66.3569.50 / 73.69 / 63.1078.83 / 86.72 / 62.19
Full78.40 / 90.77 / 66.6770.68 / 75.53 / 63.2880.61 / 85.65 / 69.98

Full improves All/New throughout, supporting joint use. Removing 3DGR lowers New by 0.40/13.06/5.39 points; MGCA 0.32/0.18/7.79, largest on ScanObjectNN and ShapeNetCore, respectively.


◆ Q3: End-to-end runtime, peak GPU memory and measurement conditions?

A3: Uni3D, 1,024 points for all methods. Entries: trainable parameters (M) / single-object forward FLOPs (G).

Table R3a: Parameters / FLOPs.

MethodModelNet40ScanObjectNNShapeNetCore
BIA8.45 / 88.488.42 / 88.488.46 / 88.48
ProtoGCD13.95 / 87.6913.93 / 87.6913.97 / 87.69
SimGCD13.95 / 87.6913.93 / 87.6913.97 / 87.69
Geo3D9.72 / 87.6810.25 / 87.699.72 / 87.68

Geo3D has 26.4%–30.4% fewer trainable parameters than ProtoGCD/SimGCD, with comparable forward FLOPs.

Timing: one RTX 4090, batch size 16, native training/evaluation procedures. Geo3D includes memory refresh, reference inference, geometry/targets, parameter updates, evaluation and checkpoint saving.

Table R3b: Epoch time (minutes).

MethodModelNet40ScanObjectNNShapeNetCore
BIA3.113.7211.49
ProtoGCD4.695.3816.53
SimGCD4.314.8615.57
Geo3D3.393.9112.45

Geo3D: 100-epoch training times 5.65/6.52/20.75 hours; peak GPU tensor memory allocated in per-epoch measurements: 3.84/3.88/4.46 GiB.

Geo3D is faster per epoch than ProtoGCD/SimGCD throughout, with improved novel discovery.

Appendix B.4, Tables 8(a)/(b), adds costs/conditions in dark green for both reviewers.

AuthorsOpenReview ↗

Response to Reviewer xHXK

Thank you again for your careful review and constructive feedback. We understand that the discussion period may be busy, and we wanted to make one final brief follow-up before it closes.

Our rebuttal directly addresses your three main questions: (1) baseline adaptation and protocol comparability, including a controlled Ward versus K-Means comparison showing consistent improvements under both clustering methods; (2) technical distinctions from AMEND, CMS, ProtoGCD, and TAN, together with ablation evidence supporting the complementary roles of 3DGR and MGCA; and (3) end-to-end computational costs, including training time and peak GPU memory measurements on an RTX 4090, accounting for the additional pipeline overhead. We have also revised the title following your suggestion.

We would be very grateful to know whether any of your original concerns remain insufficiently addressed or whether any further clarification would be helpful for your assessment. We would be happy to provide additional explanations if needed.

Thank you again for your time, consideration, and valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.