acceptodds
Under review as a conference paper at ICLR 2027

Tracing Chains of Evidence: Cross-View Test-Time Adaptation for Point Cloud Part Segmentation

Abstract

Point cloud part segmentation increasingly lifts multi-view predictions from 2D foundation models into 3D, reducing the need for costly point-level annotations. To assess the robustness of this paradigm, we establish, to the best of our knowledge, the first distribution-shift benchmark for 2D-to-3D part segmentation, spanning corruption, category, and sim-to-real shifts. Our evaluation reveals substantial performance degradation and exposes an inherent evidence-coupling challenge: multi-view predictions linked through shared 3D geometry can be complementary, ambiguous, or conflicting, hindering their reliable integration. To address this challenge, we propose **T**est-time **R**eliable **A**daptation via **C**hain-of-**E**vidence (**TRACE**), a novel TTA framework that recasts adaptation as discovering and reinforcing structural consensus within entangled evidence. Specifically, TRACE constructs a weighted evidence graph that jointly models semantic compatibility and 3D geometric consistency, organizing compatible cross-view predictions into evidence chains while routing ambiguous evidence to an explicit uncertainty state. To avoid the exponential cost of enumerating these chains, we derive an exact composition rule through mathematical induction that aggregates their support in closed form with linear complexity. Together, these designs move TTA beyond confidence-driven refinement of isolated predictions toward uncertainty-aware reasoning over structured cross-view consensus, enabling robust integration of complementary evidence. Extensive experiments across diverse domain shifts and base models demonstrate that TRACE achieves state-of-the-art performance, with an average improvement of +2.6% mIoU. Code will be publicly available.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Test-time Adaptation, Transfer Learning, Part Segmentation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_x8YtOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies test-time adaptation for 2D-to-3D point cloud part segmentation under geometric corruption, category shifts, and synthetic-to-real transfer. It introduces TRACE, which organizes predictions from multiple views into semantically compatible evidence chains, calibrates them using overlap between their 3D supports, and uses an exact closed-form composition rule to avoid enumerating the chains. An uncertainty-gated objective then adapts sift prompts while keeping the perception and lifting models frozen. Experiments with PartSLIP and PartSTAD, five TTA baselines, component ablations, and a real-world demonstration show consistent improvements across the three shift settings.

Critical strengths
  1. The topic of robustness and generalization in 2D-to-3D perception models is both timely and important, as it is a common challenge in many real-world systems, yet somewhat overlooked by the TTA community.

  2. The central idea of “chains of evidence” is novel and intuitive. By recognizing that predictions across views are coupled and may conflict, the paper explains why conventional TTA struggles in the 2D-to-3D lifting pipeline. Tracing compatible evidence chains then provides a coherent basis for adaptation.

Critical weaknesses
  1. PartSLIP and PartSTAD differ in region extraction and lifting, but both use GLIP and the same ten fixed views. Evidence with fewer valid views or a different 2D detector would help establish how broadly TRACE’s gains apply.

  2. The ablations support the value of geometric calibration and uncertainty gating, but do not include a closely matched, simpler cross-view consensus baseline using the same prompt updates and adaptation objective. It is therefore difficult to isolate the benefit of evidence-chain consolidation from geometric reweighting and selective updates.

Questions for authors
  1. How does TRACE perform with fewer valid views or a different 2D detector? This would clarify the scope of the reported gains.

  2. How reliable is the uncertainty gate when different views confidently support conflicting classes?

  3. With the same prompt parameterization, update budget, and adaptation objective, how does TRACE compare with a simple geometry-weighted cross-view consensus baseline? This would clarify the specific contribution of evidence-chain consolidation.

5 comments
AuthorsOpenReview ↗

Response to x8Yt (part I)

Thank you for your careful review and positive assessment of our work. We especially appreciate your recognition of the importance of robustness in 2D-to-3D perception and the value of reasoning jointly about cross-view evidence. Your constructive suggestions help us clarify TRACE’s applicability and strengthen its empirical evaluation. We address each concern below with additional results and methodology clarification.


W1 / Q1(a). Evidence with fewer valid views would help establish how broadly TRACE’s gains apply.

A1: Thank you for this constructive suggestion. We additionally evaluated TRACE with 6, 8, and 10 rendered views on both PartSLIP and PartSTAD under category shift. R-Table 11 shows that fewer views reduce visible-point coverage. At six views, ZeroSiam underperforms both frozen models, while TRACE improves their mIoU by 1.01 and 1.64%, respectively. TRACE consistently improves over the frozen models across all six configurations, supporting its effectiveness with fewer available views. We have included this experiment and its analysis in Section 6.4 of the revised manuscript.

R-Table 11: Effect of view count on category shift. Visible-point coverage is the percentage of points visible from at least one rendered view.

BackboneViewsVisible-point coverageFrozenZeroSiamTRACE
PartSLIP678.27%19.8919.4520.90
PartSLIP886.97%22.9723.2225.35
PartSLIP1095.16%23.0622.4425.83
PartSTAD678.27%21.2119.9622.85
PartSTAD886.97%29.5029.7432.50
PartSTAD1095.16%30.6931.4434.97

W1 / Q1(b). How does TRACE perform with a different 2D detector?

A2: Thank you for this valuable question. We replaced GLIP with Grounding DINO [1] in the PartSLIP pipeline under the six category-shift directions. R-Table 12 shows that TRACE outperforms the strongest compared baseline by +8.36% mIoU on average, with improvements in every direction. This gain exceeds that observed with the original GLIP-based PartSLIP, suggesting further potential when TRACE is paired with stronger detectors. We have included these results and analysis in Table 4 of the revised manuscript.

R-Table 12: Category-shift results on additional pipelines. PartSLIP + Grounding DINO replaces the detector with Grounding DINO.

ModelAdaptationL→HH→LC→UU→CR→II→ROverall Avg.
PartSLIP + Grounding DINOReCAP36.8625.1035.9029.4239.9027.4232.43
PartSLIP + Grounding DINOZeroSiam36.7725.2735.8929.5539.9527.4932.49
PartSLIP + Grounding DINOTRACE46.7328.1743.2740.5750.2736.1140.85
AuthorsOpenReview ↗

Response to x8Yt (part II)

Q2. How reliable is the uncertainty gate when different views confidently support conflicting classes?

A3: Thank you for this question. We assessed the gate’s reliability by examining whether points with conflicting cross-view support pass it. We measure conflict using the mean pairwise Jensen–Shannon divergence, DD, across valid visible views and classify points as high conflict when D>0.2D>0.2.

R-Table 13 shows that the gate filters out most high-conflict points, with those passing it accounting for 0.62–1.35% of all valid points. Additionally rejecting these remaining points improves mIoU by only 0.06 / 0.02 / 0.12 points, indicating a small net effect on segmentation performance. We have included this empirical validation in Appendix C.4 of the revised manuscript.

R-Table 13: Conflict statistics and additional conflict rejection. The final column adds rejection of points with D>0.2D>0.2 to the original gate.

SettingHigh-conflict points (%)High-conflict points passing the gate (%)TRACE mIoUTRACE + conflict rejection
Corruption shift17.841.1435.6035.66
Category shift9.780.6234.9734.99
Sim-to-real shift24.631.3538.5138.63

W2 / Q3. How does TRACE compare with a simple geometry-weighted cross-view consensus baseline?

A4: Thank you for this question. Following your suggestion, we compare TRACE with different cross-view fusion baselines under the same prompt parameterization, update budget, and adaptation objective: Averaging, Normalized product, and geometry weighting. R-Table 14 shows that TRACE consistently outperforms all three baselines across both backbones and all three shift types. Among the simpler baselines, geometry weighting achieves the highest average mIoU on both backbones, supporting the value of geometric reliability in combining cross-view evidence. We have included this comparison and analysis in Appendix C.5 of the revised manuscript.

R-Table 14: Comparison with cross-view fusion baselines under category shift.

BackboneFusionComplexity shiftComposition shiftShape shiftAverage
PartSLIPAveraging24.1022.9525.1624.07
PartSLIPNormalized product23.7422.9224.7623.81
PartSLIPGeometry weighting24.5423.0225.0624.21
PartSLIPTRACE25.9124.0527.5325.83
PartSTADAveraging30.9931.9333.5732.16
PartSTADNormalized product30.3131.8333.2731.80
PartSTADGeometry weighting32.1932.5334.8033.17
PartSTADTRACE34.4235.1335.3734.97

References

[1] Shilong Liu et al. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. ECCV, 2024.

Thank you for the detailed rebuttal. All my concerns have been satisfactorily addressed. The additional experiments support TRACE’s effectiveness across different view counts and detectors, while the conflict analysis and view-fusion baselines further validate the proposed design. I have decided to raise my score from 3 to 4.

Reviewer_pptJOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TRACE is a test-time adaptation method for 2D-to-3D part segmentation pipelines (PartSLIP and PartSTAD) that updates only GLIP prompt offsets on a stream of unlabeled target shapes. For each point, per-view part scores become normalized evidence weights with an explicit uncertainty state (Eq. 5); each view is down-weighted by a reliability taken from the dominant eigenvector of a view-overlap IoU matrix. The paper also builds a benchmark with corruption shift (PartNet-C, following PointCloud-C), category shift (six PartNet-Mobility splits), and sim-to-real shift (12 ScanObjectNN categories), and reports average gains of about 2.6 mIoU over five TTA baselines, averaged over three runs.

Critical strengths

The empirical picture is consistent, and the benchmark is useful. Gains hold for both base models, all six category directions, all twelve ScanObjectNN categories, and every corruption type, while several standard TTA objectives fall below the unadapted model. The view-wise versus aggregation-based analysis (Appendix C.1) and the correlation between inter-view conflict and gain (Spearman about 0.7, Fig. 5) support the paper's diagnosis that cross-view conflict is what breaks adaptation in this setting.

Critical weaknesses

The theoretical contribution is a re-derivation of Dempster-Shafer evidence combination, and that literature is not cited. With m = 1, Eq. 5 is the evidential deep-learning opinion (Sensoy et al., NeurIPS 2018), Eq. 8 is Shafer's discounting, Proposition 1 is Dempster's rule as already used for multi-view fusion in Trusted Multi-View Classification (Han et al., ICLR 2021), and Proposition 2 is the distributive law, as your own Eq. 22 shows. Contributions 2 and 3 therefore do not hold as stated, and since Table 6 never replaces the clique rule with simple fusion under the same loss, what the rule itself adds is unknown.

The size of the improvement is uncertain because the comparison may favor TRACE. All methods share one prompt-tuning setting (AdamW, lr 0.07, weight decay 0.25) that appears to be chosen for TRACE, which may explain why TENT, SAR and DUO fall below the unadapted model; TRACE's m, tau and lambda all sit at the peaks of sweeps run on the category-shift test tasks (Fig. 4); and the 1.7 to 3.5 mIoU margins are means of three runs with no spread.

Questions for authors
  1. Is there any difference between Propositions 1-2 and Dempster's rule as used in Trusted Multi-View Classification? If not, please reposition the contribution and cite this line of work.
  2. The same loss and gate, with the per-view distributions fused by averaging or by a normalized product instead of the clique rule.
  3. Baselines with per-method learning-rate selection under a protocol that does not read target labels, and a description of how m, tau, and lambda were chosen.
  4. Standard deviations over the three runs, and what varies between runs given fixed cameras and a fixed stream.
4 comments
AuthorsOpenReview ↗

Response to pptJ (part I)

Thank you for your careful review and detailed feedback. We especially appreciate your recognition of the benchmark’s value, TRACE’s consistent improvements across distribution shifts, and our analysis of how cross-view conflict undermines adaptation in this task. Your constructive suggestions help us clarify TRACE’s relationship to prior work and strengthen the fairness of our evaluation. We address each concern below with methodological clarifications and additional experiments.


W1 / Q1. Relationship to evidential learning and Dempster–Shafer combination.

A1: Thank you for this careful analysis and constructive suggestion. We have added the relevant citations and clarified the mathematical foundations in Section 4.2 and Appendix A. Propositions 1–2 now characterize the adopted DS-based computation within TRACE’s graph formulation, with the following attribution:

ComponentRelationship to prior work
Eq. (5)The evidential opinion at m=1m=1 [1].
Eq. (8)Discounting in evidence theory [2], with reliability estimated from 3D support overlap in TRACE.
Proposition 1, after normalizationDempster's combination rule [3].
Proposition 2The closed-form expansion via the distributive law.

TRACE’s contribution concerns how evidence is constructed and used for adaptation in 2D-to-3D segmentation, with the following methodological differences and empirical benefits:

  1. Geometric reliability estimation. TMC estimates uncertainty from learned class evidence. TRACE additionally estimates geometric reliability from the dominant eigenvector of a 3D support-overlap matrix.
  2. Unsupervised test-time adaptation. TMC trains evidence predictors using labeled classification data. TRACE updates only prompts on unlabeled target shapes, guided by calibrated evidence and uncertainty gating.
  3. Effects on 2D-to-3D pipelines. R-Table 5 shows that TRACE achieves +4.16% mIoU over DS fusion, whereas DS + SAR and DS + ReCAP both underperform the frozen model on average. These results support the benefit of TRACE’s complete adaptation procedure beyond directly applying DS fusion to TTA. We have included this comparison in Appendix C.3 of the revised manuscript.

R-Table 5: Comparison with DS-based fusion and adaptation variants under category shift. DS fusion changes final fusion without prompt updates.

VariantComplexityCompositionShapeAverage
Frozen model29.1530.5632.3830.69
DS fusion29.7730.4832.1730.81
DS + SAR26.7631.6230.8929.76
DS + ReCAP28.5129.8431.2629.87
TRACE34.4235.1335.3734.97

Accordingly, we reposition Contributions 2–3 around geometry-calibrated cross-view evidence and selective prompt adaptation. The graph formulation connects semantic compatibility and shared 3D support to point-level adaptation signals; Propositions 1–2 explain its correspondence to the established DS computation. We have clarified the contribution statements in abstract and introduction of the revised manuscript.


W1 / Q2. Replacing evidence combination with averaging or a normalized product.

A2: Thank you for this suggestion. We evaluated averaging and normalized-product alternatives on the category-shift benchmark. R-Table 6 shows that TRACE outperforms averaging by 1.76 / 2.81% mIoU and normalized product by 2.02 / 3.17% on PartSLIP / PartSTAD, respectively, with improvements across all three shift types. These results support the effectiveness of TRACE’s geometry-calibrated evidence combination over simpler cross-view fusion alternatives. We have included this comparison in Appendix C.5 of the revised manuscript.

R-Table 6: Comparison of evidence combination methods under category shift.

BackboneFusionComplexityCompositionShapeOverall mIoU
PartSLIPAveraging24.1022.9525.1624.07
PartSLIPNormalized product23.7422.9224.7623.81
PartSLIPTRACE evidence combination25.9124.0527.5325.83
PartSTADAveraging30.9931.9333.5732.16
PartSTADNormalized product30.3131.8333.2731.80
PartSTADTRACE evidence combination34.4235.1335.3734.97
AuthorsOpenReview ↗

Response to pptJ (part II)

W2 / Q3. Per-method learning-rate selection and selection of mm, τ\tau, and λ\lambda.

A3: Thank you for this suggestion. We select hyperparameters on 200 PartNet validation samples covering six corruption types at severity 4. The validation set is disjoint from the test set by original 3D shape and excludes target object categories. Selection uses validation mIoU, while target evaluation labels remain excluded from tuning and test-time adaptation. We have included the selection protocol and results in Section 5.1 and Appendix B.5 of the revised manuscript.

Following the optimizer configuration of PartSLIP and PartSTAD, we used AdamW with a weight decay of 0.25 and searched learning rates in {0.01,0.03,0.05,0.07,0.10}\{0.01,0.03,0.05,0.07,0.10\}. The shared learning rate of 0.07 achieved the highest average validation mIoU across the baselines in R-Table 7. R-Table 8 reports test-set sensitivity across the same learning rates, with TRACE outperforming all compared baselines at every tested learning rate.

R-Table 7: Learning-rate selection on the validation set. Bold values indicate the highest mIoU in each row.

Methodlr = 0.01lr = 0.03lr = 0.05lr = 0.07lr = 0.10
TENT14.3716.6420.3520.1915.95
SAR17.0917.1619.8320.9318.32
DUO23.5522.8623.5224.0921.08
ReCAP25.4825.8927.1227.8826.97
ZeroSiam26.0326.3427.4527.6827.10
Average21.3021.7823.6524.1521.88

R-Table 8: Learning-rate sensitivity on the category-shift test set with PartSTAD.

Methodlr = 0.01lr = 0.03lr = 0.05lr = 0.07lr = 0.10
TENT22.6221.0123.1323.7320.12
SAR24.0922.5023.7224.1821.08
DUO26.2526.7727.1026.9725.86
ReCAP30.6230.8930.9531.0030.25
ZeroSiam30.8930.8531.2931.4431.05
TRACE31.2033.5834.8634.9734.21

For method-specific hyperparameters, we used the same validation criterion and equal search budgets. R-Table 9 provides the candidate ranges and selected values for ReCAP and TRACE. TRACE’s m=0.9m=0.9, τ=0.5\tau=0.5, and λ=1.0\lambda=1.0 were fixed before examining the test-set sensitivity analysis in Fig. 4. We have documented this selection protocol and updated the discussion to reflect learning-rate sensitivity.

R-Table 9: Hyperparameter search ranges and validation-selected values.

MethodHyperparameterSearch candidatesSelected value
ReCAPThreshold0.8, 0.9, 1.00.8
ReCAPScale factor0.25, 0.5, 0.750.5
ReCAPKL-weight1.0, 1.2, 1.51.2
TRACEmm0.8, 0.9, 1.00.9
TRACEτ\tau0.25, 0.5, 0.750.5
TRACEλ\lambda0.5, 1.0, 1.51.0

Q4. Standard deviations over the three runs, and what varies between runs given fixed cameras and a fixed stream.

A4: Thank you for this question. The three runs reported in the main paper use fixed camera poses but different test-stream orders. R-Table 10 shows that TRACE’s standard deviations range from 0.32 to 0.40 mIoU points, with lower variability than ZeroSiam across all four settings. These results complement the reported mean performance gains by showing limited variation across stream orders. TRACE also outperforms the strongest compared baseline in each of the three runs, demonstrating consistent gains across different stream orders.

R-Table 10: PartSLIP results across three runs. Values are mean ± standard deviation of mIoU (%).

MethodCorruption S3Corruption S5Category shiftSim-to-real shift
ReCAP25.90 ± 0.3120.65 ± 0.2823.17 ± 0.2931.12 ± 0.34
ZeroSiam26.23 ± 0.6520.59 ± 0.4822.44 ± 0.6931.58 ± 0.50
TRACE29.06 ± 0.4022.59 ± 0.3225.83 ± 0.3433.27 ± 0.38

References

[1] Sensoy, Kaplan, and Kandemir. Evidential Deep Learning to Quantify Classification Uncertainty. NeurIPS 2018.

[2] Shafer. A Mathematical Theory of Evidence. Princeton University Press, 1976.

[3] Han, Zhang, Fu, and Zhou. Trusted Multi-View Classification. ICLR 2021.

Thank you for your response. The attribution table resolves Q1. R-Table 6 shows that replacing the combination rule with averaging or a normalized product reduces mIoU by 1.76–3.17 percentage points. Together with R-Table 5, these results support the value of combining established DS fusion with the gated objective. I recognize the empirical contribution here.

Regarding Appendix C.5, were Eq. 8, the Dirichlet objective, and the gating mechanism held fixed across the fusion variants? Please also report the fraction of points passing the gate for each variant, as I think the same threshold may select different subsets of points.

The learning-rate selection protocol resolves my concern that it might favor TRACE: the shared rate was selected based on the baselines’ average validation performance. Please limit the conclusions about negative transfer to the tested learning-rate range, and provide the validation scores used to select m, τ, and λ in R-Table 9.

Please credit the combination rule to Dempster or Shafer and cite TMC for its application in multi-view learning. In positioning the contribution, I would suggest emphasizing the combination of the fusion method and the objective; the additional gains from geometric calibration in Table 6 are relatively small.

Q4 has been answered. Please include the PartSTAD standard deviations in the final version. Also, the “Prior SOTA” line in Fig. 4 appears to indicate the frozen PartSTAD result, whereas Table 2 reports 31.44 for ZeroSiam.

Based on the above, I am raising my rating to 3, provided that the revised abstract and introduction no longer present the closed-form combination formula or the exponential-to-linear complexity reduction as new theoretical derivations.

Dear Reviewer pptJ,

Thank you for recognizing the empirical contribution of TRACE and for the helpful follow-up suggestions. We clarify the experimental setting and address the remaining points below.

Fusion variants and gate-passing rates. All variants use the same Dirichlet objective and share TRACE’s per-point DS-based gate mask, yielding identical point selection with a 73.36% passing rate. Geometric calibration in Eq. (8) is only used in TRACE. We have clarified these controls in Appendix C.5.

Learning-rate scope and hyperparameter selection. We have limited our negative-transfer conclusions to the evaluated settings and tested learning rates. The updated R-Table 9 reports validation mIoU for all 27 candidate combinations. The selected configuration, (m,τ,λ)=(0.9,0.5,1.0)(m,\tau,\lambda)=(0.9,0.5,1.0), achieves the highest validation mIoU of 28.94. These parameters were fixed before examining test-set sensitivity.

Updated R-Table 9: Validation mIoU (%) for the joint hyperparameter search. Bold indicates the best configuration.

mmτ\tauλ=0.5\lambda=0.5λ=1.0\lambda=1.0λ=1.5\lambda=1.5
0.80.2527.9728.1528.21
0.80.5028.1628.4928.50
0.80.7528.1328.2128.17
0.90.2528.2428.7228.48
0.90.5028.7928.9428.83
0.90.7528.3028.6828.64
1.00.2528.0928.2728.13
1.00.5028.1828.6328.32
1.00.7527.9528.3328.09

Attribution and contribution positioning. We have explicitly credited the combination rule to Shafer (1976) and cited TMC for its application in multi-view learning. The revised abstract and introduction emphasize the integration of established DS fusion with the gated adaptation objective. They no longer present the closed-form combination formula or the exponential-to-linear computation as new theoretical derivations.

Standard deviations and Fig. 4. We will include the complete PartSTAD results as mean ± standard deviation in the final version. Thank you for identifying the baseline inconsistency in Fig. 4; we have corrected the reference line.

Thank you once again for your constructive feedback and continued support.

Best, Authors

Reviewer_36zpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TRACE, which uses multi-view predictions and overlap between their associated 3D regions to construct an adaptation signal for updating text prompts on unlabeled test data. The authors apply it to PartSLIP and PartSTAD and evaluate segmentation performance under point-cloud corruptions, category shifts, and transfer to real scans.

Critical strengths
  1. TRACE improves both PartSLIP and PartSTAD under several types of distribution shift. With PartSTAD, it improves mean mIoU over the strongest competing method by 3.53 percentage points under category shift and 3.08 points on real scans. These results suggest that the adaptation procedure can be useful for existing segmentation systems.

  2. The method checks whether detections from different views cover overlapping regions in 3D and uses this information to reweight the evidence used for prompt updates. Without ground-truth labels, this provides a geometric basis for deciding which predictions to learn from.

Critical weaknesses
  1. The core fusion formula is the same Dempster–Shafer combination rule used in Han et al., Trusted Multi-View Classification (ICLR 2021). Equation (10), after normalization, gives this rule, and Equation (11) is its multi-view expansion. The authors describe the operation using maximum cliques and evidence chains, but sequential application of the existing rule already gives linear complexity in the number of views. This is a reformulation of an existing fusion method and should not be listed as a new technical contribution.

  2. The authors use the fused uncertainty mass to decide whether to update, but a low uncertainty mass does not mean that the pseudo-label is reliable. For example, suppose one view assigns masses of 0.8 and 0.1 to the seat and backrest, respectively, while another view assigns the opposite masses, with both assigning 0.1 to uncertainty. With both geometric reliability weights set to 1, the fused class probabilities are still 50% each, but the uncertainty mass falls to about 0.029, below the default threshold of 0.5. Despite the clear disagreement between the views, this point passes the gate, and the model selects one of the two equally supported classes as a pseudo-label and reinforces that choice.

Questions for authors
  1. What does TRACE require from the underlying segmentation method? The experiments cover only PartSLIP and PartSTAD, both of which use GLIP. Can TRACE also be applied to other projection-based 3D segmentation methods? If so, I would like to see results on additional methods, particularly those using a different 2D foundation model.

  2. The paper shows that TRACE improves projection-based segmentation methods. How do the adapted systems perform under distribution shift compared with recent methods that segment parts directly in 3D? I would like to see comparisons with comparable task definitions and supervision settings to assess whether the adapted systems are competitive in segmentation quality and robustness

4 comments
AuthorsOpenReview ↗

Response to 36zp (part I)

Thank you for your careful review and positive assessment of our work. We especially appreciate your recognition of TRACE’s usefulness for existing segmentation systems and the value of geometric evidence calibration. Your constructive suggestions help us clarify TRACE’s contribution and strengthen its empirical evaluation. We address each concern below with methodological clarifications and additional comparisons.


W1: The fusion formula follows Dempster–Shafer combination as used in TMC.

A1: Thank you for pointing out this connection. We have cited TMC and clarified the graph-view reformulation of TRACE in introduction and Section 4.2 of the revised manuscript. In 2D-to-3D segmentation, this representation connects cross-view semantic compatibility with geometric reliability, making explicit how projected regions jointly contribute to point-level adaptation signals.

TRACE’s contribution lies in how evidence is constructed and used for unsupervised adaptation, with the following methodological distinctions and empirical benefits:

  1. Geometric reliability estimation. TMC estimates uncertainty from learned class evidence. TRACE estimates geometric reliability from the dominant eigenvector of a 3D support-overlap matrix.
  2. Unsupervised test-time adaptation. TMC trains evidence predictors using labeled classification data. TRACE updates only prompts on unlabeled target shapes, guided by calibrated evidence and uncertainty gating.
  3. Effects on 2D-to-3D pipelines. R-Table 1 shows that TRACE achieves +4.16% mIoU over DS fusion, whereas DS + SAR and DS + ReCAP both underperform the frozen model on average. These results support the benefit of TRACE’s complete adaptation procedure beyond directly applying DS fusion to TTA area. We have included this comparison in Appendix. C.3.

Accordingly, we reposition Contributions 2–3 around geometry-calibrated cross-view evidence and selective prompt adaptation. The graph formulation connects semantic compatibility and shared 3D support to point-level adaptation signals; Propositions 1–2 explain its correspondence to the established DS computation. We have clarified the contribution statements in the revised manuscript.

R-Table 1: Comparison with DS-based fusion and adaptation variants under category shift. DS fusion changes final fusion without prompt updates.

VariantComplexityCompositionShapeAverage
Frozen model29.1530.5632.3830.69
DS fusion29.7730.4832.1730.81
DS + SAR26.7631.6230.8929.76
DS + ReCAP28.5129.8431.2629.87
TRACE34.4235.1335.3734.97

W2: Low uncertainty mass can admit conflicting predictions and unreliable pseudo-labels.

A2: Thank you for identifying this failure case. We measure cross-view conflict by the mean pairwise Jensen–Shannon divergence, DD, across valid visible views and use D>0.2D>0.2 for this diagnostic analysis; your example gives D≈0.49674D\approx0.49674.

R-Table 2 shows that high-conflict points passing the gate account for 0.62–1.35% of all valid points, indicating that the gate filters out most high-conflict cases. In addition, the remaining high-conflict points have only a small effect (<0.15) on performance, suggesting limited sensitivity to remaining high-conflict updates. We have clarified the limitation of uncertainty gating and added this empirical validation in Appendix C.4 of the revised manuscript.

R-Table 2: Conflict statistics and additional conflict rejection. The final column adds rejection of points with D>0.2D>0.2 to the original gate.

SettingHigh-conflict points (%)High-conflict points passing the gate (%)TRACE mIoUTRACE + conflict rejection
Corruption shift17.841.1435.6035.66
Category shift9.780.6234.9734.99
Sim-to-real shift24.631.3538.5138.63
AuthorsOpenReview ↗

Response to 36zp (part II)

Q1: Can TRACE apply to other projection-based methods, particularly with a different 2D foundation model?

A3: Thank you for this constructive question. TRACE is model-agnostic within projection-based pipelines: it only uses projection correspondences, semantic scores and a differentiable parameter interface. These components are common in projection-based pipelines, allowing TRACE to extend beyond a specific detector or lifting architecture.

Following your suggestion, we further evaluated TRACE on PartSLIP++ [1] and PartSLIP with GLIP replaced by Grounding DINO[2]. R-Table 3 shows gains over the strongest compared baseline of 4.52 and 8.36 mIoU, respectively, with improvements in every transfer direction. These results support TRACE’s applicability to additional pipelines and different 2D foundation models. We have included these results in Table 4 of the revised manuscript.

R-Table 3: Additional segmentation systems under category shift. mIoU in %. L/H: low/high complexity; C/U: common/uncommon composition; R/I: regular/irregular shape.

SystemAdaptationL→HH→LC→UU→CR→II→ROverall Avg.
PartSLIP++ReCAP30.8223.1231.1922.6533.5821.9427.22
PartSLIP++ZeroSiam31.0323.2931.5722.8133.7922.2727.46
PartSLIP++TRACE36.1825.0835.4328.8939.3126.9931.98
PartSLIP + Grounding DINOReCAP36.8625.1035.9029.4239.9027.4232.43
PartSLIP + Grounding DINOZeroSiam36.7725.2735.8929.5539.9527.4932.49
PartSLIP + Grounding DINOTRACE46.7328.1743.2740.5750.2736.1140.85

Q2: How do TRACE-adapted systems compare with methods that segment parts directly in 3D?

A4: Thank you for this valuable suggestion. We evaluated Find3D (ICCV 2025) [3] and PatchAlign3D (CVPR 2026) [4] on our benchmark. Both methods predict part labels directly in 3D. For the system-level comparison, we use each method’s original pretrained model and apply the same task-training protocol (data and supervision) across methods. Test conditions are fully aligned, including the part vocabulary, label mapping, and mIoU computation protocol. For every method, training and test instances are disjoint.

R-Table 4 shows that PartSTAD + TRACE exceeds PatchAlign3D by 8.76 / 3.47 / 2.28 / 6.78 mIoU. PartSLIP + TRACE also exceeds both methods under corruption and on real scans. These results demonstrate that TRACE-adapted systems can improve projection-based pipelines to outperform 3D-native models under distribution shifts. We have included the comparison with evaluation details in Table 5 of the revised manuscript.

R-Table 4: Comparison with methods performing part prediction directly in 3D. mIoU in %.

MethodCorruption S3Corruption S5Category shiftSim-to-real shift
Find3D22.0718.7921.9825.23
PatchAlign3D26.8421.7732.6931.73
PartSLIP + TRACE29.0622.5925.8333.27
PartSTAD + TRACE35.6025.2434.9738.51

References

[1] Yuchen Zhou et al. PartSLIP++: Enhancing Low-Shot 3D Part Segmentation via Multi-View Instance Segmentation and Maximum Likelihood Estimation. arXiv:2312.03015, 2023.

[2] Shilong Liu et al. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. ECCV, 2024.

[3] Ziqi Ma, Yisong Yue, and Georgia Gkioxari. Find Any Part in 3D. ICCV, 2025.

[4] Souhail Hadgi et al. PatchAlign3D: Local Feature Alignment for Dense 3D Shape Understanding. CVPR, 2026.

Thanks for the rebuttal and the additional results. I'm satisfied with the authors' response and will maintain my original score.

AuthorsOpenReview ↗

Thank you for your feedback

Dear Reviewer 36zp,

Thank you for your follow-up and continued support. We are glad that our rebuttal has satisfactorily addressed your concerns.

Your suggestions helped us clarify TRACE’s relationship to Dempster–Shafer fusion, evaluate the effects of uncertainty gating, and strengthen our evaluation through additional segmentation pipelines, a different 2D detector, and comparisons with direct 3D methods. We have incorporated these clarifications and results into the revised manuscript.

If you feel these additions strengthen the work beyond your initial assessment, we would greatly appreciate any reconsideration of the overall rating.

Thank you again for your constructive feedback and careful review.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.