acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty-Aware Multi-Prototype Modeling and Peak-Preserving Calibration for Out-of-Distribution 3D Occupancy Prediction

Abstract

Reliable out-of-distribution (OOD) 3D occupancy prediction requires distinguishing genuinely unfamiliar objects from rare but valid semantic patterns. However, long-tailed supervision and intra-class variation can leave rare known voxels weakly supported by the learned feature distribution, causing existing OOD detectors to conflate semantic rarity with distributional novelty and produce fragmented anomaly responses. To address this issue, we present UMPOcc, a plug-and-play framework for rare-vs-unknown disambiguation through complementary distributional and spatial reasoning. Its Uncertainty-Aware Multi-Prototype Occupancy Field (UMPOF) represents each semantic class with multiple modes characterized by empirical priors and feature dispersion, yielding more reliable source-domain support for rare known patterns. Peak-Preserving Spatial Calibration (PPSC) further exploits local occupancy support to reduce spatially inconsistent low-confidence responses while preserving high-confidence anomaly peaks. Extensive experiments across five benchmarks show strong occupancy and OOD performance, with consistent gains on tail classes and effective generalization to real-world scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Prototype Modeling, Out-of-Distribution, Occupancy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_qDViOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UMPOcc for out-of-distribution detection in 3D semantic occupancy prediction. It extends prototype-guided occupancy refinement with class-conditional mixtures of prototypes, incorporating empirical mode priors and feature dispersion. It also introduces Peak-Preserving Spatial Calibration (PPSC), which combines local anomaly responses with predicted occupancy support. Experiments evaluate semantic occupancy and OOD detection across multiple driving benchmarks using SGN and VoxDet backbones.

Critical strengths
  1. The paper addresses a meaningful problem: distinguishing rare known semantic patterns from genuinely unknown objects. The framework can be integrated with existing occupancy models, and the evaluation covers multiple datasets and two backbones.

  2. The experimental analysis distinguishes representation refinement from score calibration and includes fixed-model comparisons across three OOD benchmarks. These comparisons provide credible evidence that PPSC improves the inherited anomaly scores.

Critical weaknesses

W1. The methodological novelty relative to existing prototype-based approaches is insufficiently established. ProOOD [1] already combines prototype-guided feature refinement, tail mining, and local/global prototype-based OOD scoring. Mixture-of-prototype modeling for OOD detection is also established in PALM [2]. UMPOcc introduces adaptations through mode priors, dispersion estimates, and occupancy-specific refinement, but their advantages over simpler extensions remain unclear. A matched comparison with a straightforward multi-prototype version of ProOOD, and an appropriately adapted mixture-prototype baseline, would help establish what is distinctive and necessary.

W2. The independent contribution of multi-prototype modeling remains weakly supported. In the matched VAA-KITTI ablation, UMPOF alone changes AUROC from 64.50 to 64.59, whereas PPSC alone raises it to 89.24. Adding the auxiliary density signal after PPSC also provides no consistent improvement. These results do not invalidate UMPOF’s role in semantic refinement, but they provide limited support for its proposed contribution to rare-versus-unknown separation. The offline bank-capacity experiment evaluates a frozen model and an auxiliary density score, so it does not establish the benefit of multiple modes during representation learning. A single-mode versus multi-mode training comparison with identical adaptation budgets, refinement branches, and losses is needed.

W3. The necessity of PPSC’s individual mechanisms is not demonstrated. PPSC combines local averaging, peak preservation, occupancy-weighted aggregation, support-dependent interpolation, and suppression of voxels predicted as empty. The reported comparisons establish that the combined operator is useful, but do not identify which mechanisms account for the gains. Controlled alternatives such as occupancy masking alone, local smoothing with the same neighborhood, and smoothing with peak preservation would clarify whether the full formulation is necessary. The dependence on predicted occupancy also warrants analysis, since an unknown object incorrectly classified as empty may have its anomaly response suppressed.

W4. Improved global discrimination does not consistently translate into stronger anomaly localization. Against ProOOD, VAA-KITTI-360 AUROC increases from 46.91 to 52.40, but tolerance-aware AuPRC at 1.2 m decreases from 57.50 to 51.09. On VAA-STU, AuPRC at 1.0 m decreases from 14.54 to 11.72 despite the substantial AUROC improvement. These comparisons should be distinguished from the fixed-adapted-model experiments, where adding PPSC does improve the inherited scores. The evidence therefore supports PPSC as a useful calibration component more strongly than it supports consistently superior end-to-end OOD localization. Failure analysis is needed to explain these differences.

[1] Zhang et al. ProOOD: Prototype-Guided Out-of-Distribution 3D Occupancy Prediction. CVPR, 2026.

[2] Lu et al. Learning with Mixture of Prototypes for Out-of-Distribution Detection. ICLR, 2024.

Questions for authors
  1. With adaptation budgets and all other branches fixed, how much does multi-mode training improve upon a single-prototype formulation?

  2. How does PPSC compare with occupancy masking and simpler spatial filters under identical neighborhoods and evaluation protocols?

  3. Does UMPOcc reduce false OOD alarms on rare known classes at matched anomaly recall?

  4. What explains the localization regressions relative to ProOOD on VAA-KITTI-360 and parts of VAA-STU?

1 comment

Response to Reviewer qDVi

We thank the reviewer for recognizing the importance of rare-vs-unknown disambiguation and the fixed-model evidence for PPSC. We have sharpened the novelty claims and added controlled comparisons to distinguish prototype multiplicity, spatial correction, and localization limitations.

W1–W2 & Q1. Novelty relative to ProOOD and PALM

We agree that multi-prototype learning itself is not new. ProOOD provides prototype-guided occupancy refinement; PALM investigates mixture prototypes for OOD detection. Our contribution lies in the occupancy-specific combination of class-conditional mode support, masked-voxel refinement, tail-aware supervision, and a separate spatial anomaly-score operator. Mode priors and dispersion characterize auxiliary source-density evidence; they are not credited with PPSC's large AUROC gain.

We compared direct two-mode ProOOD [1] and voxel-adapted PALM-style [2] baselines with UMPOF using the same SGN initialization, eight-epoch training budget, and evaluation protocol, with PPSC disabled. The M=1/M=2M=1/M=2 UMPOF pair also holds its losses and refinement branches fixed.

RepresentationmIoU ↑Tail mIoU ↑AUROC ↑
ProOOD, naive M=2M=211.932.7964.52
PALM-style M=2M=211.952.8264.56
UMPOF M=1M=111.922.8064.55
UMPOF M=2M=211.972.8464.59

The advantage from extra modes is modest. Table 4 establishes a different main result: PPSC alone changes AUROC from 64.50 to 89.24, compared with 64.59 for UMPOF alone. We have separated representation refinement from spatial scoring and avoid claiming that prototype multiplicity explains the primary OOD gains.

[1] Zhang et al. ProOOD: Prototype-Guided Out-of-Distribution 3D Occupancy Prediction. CVPR, 2026.

[2] Lu et al. Learning with Mixture of Prototypes for Out-of-Distribution Detection. ICLR, 2024.

Q3. False alarms on rare known classes

We compare the false-positive rate on ground-truth occupied tail voxels at the same 95% unknown recall, using score-specific thresholds:

ScorerM=1M=1 Tail FPR ↓M=2M=2 Tail FPR ↓
EchoOOD58%56%
EchoOOD + PPSC29%28%

This is a direct test of rare-known confusion and supports only a small incremental role for multiple prototypes.

W3 & Q2. Is the full PPSC design necessary?

All spatial operators use the same frozen adapted model, EchoOOD scores, 7×7×77\times7\times7 neighborhood, empty-voxel masking, and evaluation protocol. We do not count reapplication of the same hard mask as an improvement.

Operator[email protected] ↑AUROC ↑
EchoOOD (common mask)55.664.5
Gaussian + Mask70.486.0
Occupancy-weighted Mean + Mask72.287.1
Peak-preserving Mean + Mask72.887.8
PPSC w/o occupancy support73.188.2
PPSC w/o peak preservation72.187.0
PPSC w/o support gating73.488.5
Full PPSC74.489.6

These comparisons isolate generic smoothing from occupancy-based support, peak preservation, and support-dependent interpolation. PPSC's final hard empty-voxel rule remains a limitation: unknown objects incorrectly predicted as free space can receive suppressed anomaly scores.

W4 & Q4. Higher AUROC but weaker tolerance-aware localization

We agree that global discrimination and spatial localization are distinct. Crucially, the localization deficit against ProOOD is already present in the adapted model's raw EchoOOD score, before PPSC:

Benchmark / metricProOODAdapted EchoOOD+ PPSC
VAA-KITTI-360 [email protected] ↑57.5049.8651.09
VAA-STU [email protected] ↑14.549.6011.72

Thus, on a fixed model, PPSC improves localization by +1.23/+2.12, but does not fully close the end-to-end gap to ProOOD. Meanwhile, AUROC rises from 46.91 to 52.40 (KITTI-360) and 62.11 to 77.15 (STU) in the end-to-end comparisons. We have explicitly distinguished these protocols and acknowledged failures involving small, missed, or predicted-empty unknown regions; the existing metrics do not identify a single causal failure mechanism.

We believe these matched baselines, mechanism ablations, and transparent localization analysis clarify which improvements the evidence supports. We respectfully hope they address the reviewer's soundness concerns and encourage reconsideration of the rating.

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

The authors point out that existing out-of-distribution detectors for 3D occupancy prediction are unable to distinguish between rare instances of known classes and genuinely new classes. To overcome this challenge, two methods are proposed: (1) UMPOF which uses multiple learnable prototype features per class (rather than single prototype in exisitng methods), which helps covering the rare but known class instances, and (2) PPSC, which uses spatial reasoning to combine fragmented OOD detections using neighborhood anomaly prediction information and peak preservation strategy into a more spatially localized prediction. In addition to the usual occupancy loss, a contrastive objective is used to learn good prototypes, and an auxiliary classification loss is used to learn the tail classes better. This combined method is called UMPOcc. Experiments are conducted on several common benchmarks, showing improved OOD detection performance.

Critical strengths
  1. The overall ideas in the paper are easy to understand and well-reasoned. It is easy to see how both the main ideas can help in improving performance.

  2. Experiments appear to be done well, and the authors have done a good job showing the ablation study and different metrics (mIoU vs. Tail mIoU), AuPRC vs. AuROC etc. to better illustrate the complete picture.

  3. PPSC appears to be a simple and intuitive method — a form of smoothing that preserves peaks in the anomaly estimate — that can be easily added on top of any existing OOD detector for this application. It works entirely at test-time and the ablation results in Table 4 clearly show that PPSC has a much larger impact on how well the proposed method works.

Critical weaknesses
  1. While having multiple prototypes per class makes sense, there seem to be many other ad-hoc decisions that are not clear from the paper, and the writing has to be improved in many places. For example, even after reading the paper, I could not understand the role of the prior and dispersion or what they were. The exact definitions are given in the Appendix. It would be good to have this information in the main paper.

  2. Improvements in occupancy prediction seem underwhelming to me. And related to this, going back to Table 4, the main contribution seems to be PPSC. Tail mIoU does go up by at most ~2 percentage points for one of the datasets, but I am not sure how significant this is. So, while I think L_{tail} was an intuitive and interesting addition, it does not appear to have made much of a difference. I request the authors to correct me if I am wrong here.

  3. Figure 3 shows the qualitative comparisons. But, again, I am not able to understand how to come to the conclusion that the proposed method is an improvement over the baseline, as the pictures are not easily showing this, as far as I can see.

  4. The authors mention call the method Peak-Preserving Spatial “Calibration”. I think this can be misleading as “calibration” has an accepted technical meaning. Going in, I was expecting PPSC to be about making the predicted probabilities meaningful, rather than smoothing the predictions based on local prediction variations. That the model does not actually provide calibrated anomaly scores would be itself a weakness.

  5. PPSC is shown to work well, but here, it is important to have some simpler baselines to show that more standard smoothing approaches, such as Gaussian smoothing or median filtering, do not work as well. Otherwise, this contribution may not be significant. The authors have not shown this.

Questions for authors

It would be great if the authors can address the weaknesses I have mentioned. 


Overall, the paper is about a very specific application with some specific components, particularly the PPSC part. It would be good if the authors can present a more convincing argument about the importance of this problem and its relevance to the community in the introduction. At the same time, do the authors think that some general versions of these ideas could be applicable more broadly.

2 comments

Response to Reviewer cXMa

We thank the reviewer for recognizing the motivation, multi-dataset evaluation, and practical value of PPSC. We have addressed the remaining concerns through clearer methodological explanations, targeted loss and spatial-operator ablations, improved visualizations, and more precise terminology.

W1. Priors, dispersion, and mode assignment

For class kk and mode mm, πk,m\pi_{k,m} is the fraction of source features assigned to the mode (its prevalence), while νk,m\nu_{k,m} is their mean cosine distance from its center (its intra-mode variability). These terms weight mode evidence and scale its distance penalty in Eq. (3). Eq. (5)'s ri,mr_{i,m} denotes a soft, class-conditioned assignment weight for online prototype updates. We have clarified the distinction between the offline source-domain density bank and online representation refinement in Sec. 3.3.

W2. Does the tail loss explain the occupancy improvement?

We agree that the observed occupancy gains are modest and that the contribution of LtailL_{\mathrm{tail}} must be isolated. We retain the tail-refinement branch, initialization, prototype number, optimizer, training budget, and all other losses, changing only the auxiliary tail-loss weight:

Table R1. Tail-loss ablation.

Dataset / backboneLtailL_{\mathrm{tail}}mIoU ↑Tail mIoU ↑
VAA-KITTI / SGNOff11.932.79
VAA-KITTI / SGNOn11.972.84
SSCBench-KITTI-360 / VoxDetOff21.9113.25
SSCBench-KITTI-360 / VoxDetOn22.0313.53

The isolated gain is small, consistent with an auxiliary rather than dominant role for LtailL_{\mathrm{tail}}. The larger end-to-end comparison with ProOOD-V on SSCBench-KITTI-360 (11.52→13.53 tail mIoU) reflects the overall representation framework, not the tail loss alone. We have made this attribution explicit.

W3. Qualitative comparisons

Fig. 3 now contains larger, higher-resolution crops, more legible anomaly regions and class boundaries, and consistent anomaly-score color scales. The discussion identifies which rare-known structures are better preserved and where unknown-object responses become more spatially concentrated.

W4. Meaning of “calibration”

We agree that conventional calibration concerns the relationship between predicted confidence and empirical accuracy. PPSC modifies spatial anomaly scores, not semantic logits or predictive probabilities. We have therefore renamed it Peak-Preserving Spatial Correction (PPSC) throughout, preserving the acronym. Appendix Table 17 reports only minor semantic ECE changes (51.71→51.53) and slightly worse geometric ECE (4.64→4.66); we do not present PPSC as probability calibration.

W5. Is PPSC more than conventional smoothing?

We evaluated all operators on the same frozen adapted SGN checkpoint and EchoOOD scores, using a common 7×7×77\times7\times7 neighborhood and identical empty-voxel masking. The original EchoOOD result already follows the shared masking convention; no improvement is attributed merely to reapplying it.

Table R2. Controlled spatial alternatives on VAA-KITTI.

Operator[email protected] ↑AUROC ↑
EchoOOD (common mask)55.664.5
Mean + Mask69.985.6
Gaussian + Mask70.486.0
Median + Mask69.584.9
Occupancy-weighted Mean + Mask72.287.1
Peak-preserving Mean + Mask72.887.8
Full PPSC74.489.6

Unlike indiscriminate filtering, PPSC uses occupancy-conditioned neighborhood evidence and peak retention to reduce unsupported responses without uniformly blurring compact anomalies. The controlled comparison tests this specific advantage rather than relying on an unmatched post-processing baseline.

Broader significance and generalization

The rare-known/unknown distinction matters for voxel-level spatial reasoning: scarce known objects should not trigger false alarms, while small unknown obstacles should remain detectable. PPSC uses a dense anomaly map and an occupancy-support field; similar mechanisms may be applicable to other dense or volumetric anomaly tasks. We describe this as potential generalization, not as experimentally verified performance outside the studied 3D occupancy benchmarks.

We believe the isolated loss study, fair smoothing controls, corrected terminology, and clearer visual evidence address the reviewer's concerns. We respectfully hope they encourage reconsideration of the rating.

Reviewer_cXMaOpenReview ↗

Response to authors

I thank the authors for the replies. The added evidence to show that PPSC is useful helps, but it also shows that simpler methods can already help quite a bit. Overall, my assessment remains the same. The methods are highly specific to this application, the writing needs improvement, and one of the proposed components, L_tail does not seem to have any signficant gains for me to increase my score.

Reviewer_V3BcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes UMPOcc for distinguishing rare known classes from unknown objects in 3D occupancy prediction. It combines multi-prototype feature refinement with occupancy-aware spatial calibration of anomaly scores. Experiments across five benchmarks and two occupancy backbones show improved occupancy accuracy and OOD AUROC, with most OOD gains coming from spatial calibration.

Critical strengths

The paper addresses an important problem: distinguishing rare known objects from genuinely unknown objects in 3D occupancy prediction.

PPSC delivers substantial empirical gains. In the fixed-model comparison on VAA-KITTI, it improves AUROC from 64.59 to 89.63, with improvements over the corresponding local baselines on the other two OOD benchmarks.

The framework integrates with existing occupancy networks and introduces little additional inference overhead under the reported settings.

Critical weaknesses
  1. The central claim of rare-versus-unknown disambiguation is not convincingly established. In Table 4, UMPOF alone changes AUROC only from 64.50 to 64.59, whereas PPSC alone reaches 89.24. The main OOD scoring path also excludes the proposed distributional energy. Without direct measurements of rare-known false positives versus unknown-object recall, the results demonstrate spatial postprocessing gains more clearly than the claimed benefit of multi-modal distributional modeling. This substantially weakens the evidence for the paper’s central scientific contribution.

  2. The reported tail-class gains are concentrated rather than broadly distributed. On VAA-KITTI with VoxDet, tail mIoU improves only from 3.96 to 4.15 over ProOOD-V, largely driven by other-vehicle increasing from 3.57 to 9.22. Meanwhile, bicycle, motorcycle, and person deteriorate; excluding other-vehicle, the remaining ten tail classes’ average IoU decreases from approximately 4.00 to 3.64. Although improvements need not occur in every category, this concentration leaves insufficient evidence that multi-prototype modeling reliably preserves diverse rare-known patterns. This weakens the empirical support for a central motivation of the method.

Questions for authors

Q1: Can you provide direct evidence that UMPOF improves the separation between rare known objects and unknown objects, beyond the gains from PPSC?

Q2: Why are the tail-class gains on VAA-KITTI with VoxDet mainly driven by other-vehicle, while bicycle, motorcycle, and person deteriorate? Are these trends consistent across repeated runs?

Q3: How does PPSC compare with simple spatial smoothing or occupancy-weighted averaging under the same settings?

Q4: Were the ProOOD baselines given the same additional fine-tuning budget as UMPOcc?

2 comments

Response to Reviewer V3Bc

We thank the reviewer for recognizing the importance of rare-known/unknown disambiguation, the gains from PPSC, and the low computational overhead. We have strengthened the evaluation with a direct rare-known false-alarm metric, class-wise repeatability analysis, controlled spatial baselines, and an equal-budget comparison.

W1 & Q1. Rare-known versus unknown discrimination

AUROC alone cannot establish whether rare known objects are mistaken for novel ones. We therefore measure the false-positive rate on ground-truth occupied tail-class voxels at 95% unknown-voxel recall, using the same evaluation set and score-specific thresholds.

Model / anomaly scoreTail FPR@95% TPR ↓
UMPOF M=1M=1 / EchoOOD58%
UMPOF M=2M=2 / EchoOOD56%
UMPOF M=1M=1 / EchoOOD + PPSC29%
UMPOF M=2M=2 / EchoOOD + PPSC28%

The M=1/M=2M=1/M=2 pair holds initialization, eight-epoch adaptation, losses, refinement branches, and OOD scorer fixed. These comparisons characterize the modest contribution of prototype multiplicity; most OOD-score improvement arises from PPSC (Table 4: AUROC 64.50→89.24 even without UMPOF). The main score uses EchoOOD + PPSC rather than auxiliary density fusion, which is not consistently beneficial (Appendix Table 9).

W2 & Q2. Class concentration and repeatability

We agree that VAA-KITTI/VoxDet tail improvements are uneven. In the original Table 1, other-vehicle rises 3.57→9.22 IoU, whereas bicycle 2.89→0.64, motorcycle 3.78→2.59, and person 2.62→1.77 decline. Aggregate tail mIoU increases only 3.96→4.15; excluding other-vehicle, the mean decreases. Thus, the improvement is not uniform across rare classes.

We examined the same three seeds for both methods; the repeatability summary is:

ClassProOOD-V mean±stdUMPOcc mean±std
Other-vehicle3.6±0.69.0±0.9
Bicycle2.8±0.50.8±0.3
Motorcycle3.7±0.52.7±0.5
Person2.6±0.41.8±0.4
Tail mIoU4.0±0.24.2±0.2

The revised discussion acknowledges both the concentration of gains and failures on extremely sparse categories; it does not claim consistent improvement for every tail class.

Q3. Fair comparison against simple spatial operators

All operators use the same frozen UMPOF-adapted SGN model, EchoOOD scores, 7×7×77\times7\times7 neighborhood, and empty-voxel handling. The original EchoOOD row is evaluated under the common masking convention; we do not credit PPSC for merely reapplying that mask.

Operator[email protected] ↑AUROC ↑
EchoOOD (common mask)55.664.5
Mean + Mask69.985.6
Gaussian + Mask70.486.0
Median + Mask69.584.9
Occupancy-weighted Mean + Mask72.287.1
Peak-preserving Mean + Mask72.887.8
PPSC74.489.6

This isolates the benefit of occupancy-conditioned aggregation and peak preservation beyond ordinary smoothing under the same mask.

Q4. Equal additional adaptation budget

The published ProOOD rows use the released checkpoints, whereas UMPOF receives eight adaptation epochs. We therefore performed a VAA-KITTI/SGN-specific matched-budget comparison, continuing ProOOD for the same eight epochs with identical optimization settings:

MethodmIoU ↑Tail mIoU ↑AUROC ↑
Original ProOOD checkpoint11.862.7664.50
ProOOD + 8 epochs11.902.7864.60
ProOOD + 8 epochs + PPSC†11.902.7889.25
UMPOcc + 8 epochs + PPSC11.972.8489.63

The controlled comparison separates improvements due to continued training from those due to spatial correction. PPSC can also be applied directly to ProOOD; UMPOF's incremental occupancy benefit should not be conflated with PPSC's larger AUROC contribution.

We believe the direct false-alarm metric, fair spatial baselines, class-level failure analysis, and SGN matched-budget comparison address the central soundness concerns. We respectfully hope this evidence supports reconsideration of the rating.

Thank you for the detailed responses and additional experiments. While the new results provide useful clarification, the gains from the multi-prototype design remain modest and do not sufficiently address my concern about its effectiveness. I will therefore maintain my rating.

Reviewer_U6gaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on 3D semantic occupancy prediction from an RGB image. The main contribution is a new method that tackles out-of-distribution classes. The authors design a module using multiple prototypes to model each known classes and then propose to calibrate anomaly scores using local occupancy and spatial evidence. Experiments on multiple occupancy and OOD benchmarks shows the effectiveness of the proposed method.

Critical strengths
  • The propose of OOD prediction for 3D occupancy prediction is an important problem and the main idea is sound.

  • The results show improvements in OOD detection, particularly on VAA-KITTI. The experiments also cover different backbones and datasets, providing useful evidence of generalization.

Critical weaknesses
  • I main concern is about the quality of clarity and presentation. I found the paper difficult to follow. Some terms, such as "empirical priors" and "feature dispersion," are frequently used without sufficient explanation of their roles or differences. Although their mathematical definitions are provided, more intuitive explanations would help audience understand the method. The method section is also quite dense, with many equations and symbols introduced without clear descriptions. For example, (r_{i,m}) in Eq. (5) could be explicitly described as the soft assignment weight of feature (i) to mode (m).

  • While the proposed modules are sound, I find the overall framework somewhat complicated. The ablation results in Table 4 suggest that most of the OOD improvement comes from PPSC, while UMPOF provides limited gains. This raises the question of whether the additional complexity is justified. Although computational overhead is reported in Appendix D.5, a clearer discussion of the practical benefits and costs of each component would be helpful.

  • The qualitative results are limited with low resolution. The highlighted regions in Figure 3 are hard to inspect.

  • The titles in the paper and OpenReview system are not aligned. Though this is not a weakness, the authors should be careful when doing the submission.

Questions for authors

Since UMPOF contributes little in Table 4 and Appendix Table 9, what evidence supports the need for multi-prototype modeling?

2 comments

Response to Reviewer U6ga

We thank the reviewer for recognizing the importance of OOD 3D occupancy prediction and the evaluation across multiple datasets and backbones. We have addressed the methodological, empirical, and presentation concerns through clearer explanations and a controlled prototype comparison.

W1. Empirical priors, dispersion, and notation

For class kk and mode mm, the empirical prior πk,m\pi_{k,m} measures the fraction of source features assigned to that mode (its prevalence), and dispersion νk,m\nu_{k,m} is their average cosine distance to the mode center (its variability). In Eq. (3), the prior weights mode evidence while dispersion scales the feature-distance penalty. In Eq. (5), ri,mr_{i,m} denotes the soft assignment of feature ii to mode mm within its ground-truth class. We have incorporated these interpretations and defined the symbols at first use in Sec. 3.3.

W2 & Q1. Does multi-prototype modeling provide an independent benefit?

We agree that PPSC contributes most of the OOD improvement. In Table 4, PPSC alone increases AUROC from 64.50 to 89.24, whereas UMPOF alone increases it to 64.59. The two modules have different purposes: UMPOF refines semantic occupancy representations; PPSC corrects spatial anomaly responses. We have explicitly separated these contributions rather than attributing the large AUROC gain to multiple prototypes.

To isolate prototype multiplicity, we conducted a matched M=1M=1 versus M=2M=2 experiment on VAA-KITTI/SGN. Both variants use the same ProOOD checkpoint, eight adaptation epochs, optimizer, warm-up, losses, feature-refinement branches, and EchoOOD scoring; PPSC is disabled. Only the number of modes changes.

Table R1. Prototype-count ablation (VAA-KITTI/SGN).

UMPOF modesmIoU ↑Tail mIoU ↑[email protected] ↑AUROC ↑
M=1M=111.922.8055.6064.55
M=2M=211.972.8455.6964.59

This comparison indicates modest gains from multiple modes, particularly for semantic and tail-class occupancy. A single center can collapse distinct intra-class appearances; multiple centers provide more flexible known-class support. We do not claim a large independent OOD improvement. Appendix Table 9 also shows that explicitly fusing UMPOF density into the final anomaly score offers no consistent gain, supporting the adopted EchoOOD + PPSC scoring path.

The end-to-end representation framework also improves tail mIoU over ProOOD-V from 6.34 to 7.28 on SemanticKITTI and 11.52 to 13.53 on SSCBench-KITTI-360 (Tables 1 and 3). These broader results support the framework but do not, by themselves, isolate M>1M>1; Table R1 addresses that distinction.

W2. Computational cost versus benefit

The measured overhead is limited (Appendix Table 18): on SGN, throughput changes from 2.09 to 2.08 FPS, peak memory from 4.04 to 4.05 GiB, and network parameters remain 28.40M. On VoxDet, throughput changes from 5.99 to 5.96 FPS and memory from 1.55 to 1.79 GiB. PPSC has no trainable parameters. We have clarified that the incremental multi-mode gain is modest, while the complete framework provides substantial OOD-score improvements at low additional inference cost.

W3. Qualitative presentation

We have improved Fig. 3 with higher-resolution crops, larger highlighted regions, clearer annotations, and consistent anomaly-score color scales. The accompanying text identifies the visible differences in rare-class reconstruction and localized anomaly responses.

W4. Title consistency

We have aligned the manuscript title and OpenReview metadata.

The controlled comparison clarifies UMPOF's limited but distinct representation benefit, while the existing ablations establish PPSC as the principal source of OOD-score gains. Together with the measured overhead and clearer presentation, we believe these results address the reviewer's main concerns and respectfully hope they encourage reconsideration of the rating.

Reviewer_U6gaOpenReview ↗

Keep the score

Thank you to the authors for the rebuttal, which has helped improve the paper. But my overall evaluation of the work remains unchanged. I encourage the authors to incorporate these additional results (along with those addressing the concerns raised by other reviewers) into the final version. I also encourage the authors to further improve the presentation and the overall quality.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.