acceptodds
Withdrawn from ICLR 2027

Instance-Centric Unsupervised Monocular 3D Occupancy Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Unsupervised monocular 3D occupancy prediction aims to recover dense scene geometry from a single image without 3D dense occupancy annotations, but often produces blurred object boundaries and trailing artifacts under sparse and indirect image-based supervision. We attribute these limitations to the prevailing scene-level representation, which learns unified geometry without explicitly distinguishing individual object geometries, thereby overlooking the 3D shape modeling of individual objects. To address this issue, we propose ICOcc, an instance-centric unsupervised occupancy prediction framework, which reformulates occupancy learning from scene-level reconstruction to instance-centric prediction. Specifically, an instance-centric categorical occupancy formulation directly separates the entangled object geometries, without the need for a unified scene geometry. Based on this formulation, an instance decoupling network is introduced, where learnable queries are employed to capture shape priors and decouple object geometries from scene-level visual features. To supervise the decoupled outputs, an instance-anchored learning strategy further provides stronger constraints through multi-view supervision across dynamically selected views. Extensive experiments on KITTI-360 and CarlaOcc show that ICOcc improves occupancy prediction, achieving a 10.2% relative gain in invisible-empty recall on KITTI-360, while producing clearer object boundaries and fewer trailing artifacts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Occupancy Prediction, Autonomous Driving, Neural Radiance Field, Computer Vision, Unsupervised Learning

Reviews

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

This paper proposes ICOcc, an instance-centric framework for unsupervised monocular 3D occupancy prediction. Instead of reconstructing a unified scene-level occupancy field, ICOcc decomposes occupancy into multiple latent slots to model object-level geometries separately. The framework further introduces a query-based Instance Decoupling Network and an Instance-Anchored Learning Strategy with multi-view pseudo supervision. Experiments on KITTI-360 and CarlaOcc show improved occupancy prediction.

Critical strengths
  1. The instance-centric occupancy formulation is well motivated and provides a meaningful alternative to conventional scene-level reconstruction.

  2. The experimental results are consistent with the paper’s motivation, particularly in improving object boundaries and reducing trailing artifacts.

Critical weaknesses
  1. The experiments do not fully isolate the benefit of the instance-centric formulation from the additional pseudo segmentation, depth, surface, and free-space supervision. A controlled scene-centric baseline with the same supervision would make the contribution clearer.

  2. The evidence that learnable queries truly capture stable instance-level 3D shape priors remains indirect. The paper also lacks analysis of query capacity, such as sensitivity to the number of queries or behavior in crowded scenes.

Questions for authors
  1. How much of the performance gain comes from the instance-centric formulation itself, rather than the additional pseudo segmentation, depth, and geometric supervision? Could the authors provide a scene-centric baseline using the same supervision signals?

  2. Could the authors provide more direct evidence that the queries learn instance-level 3D shape priors?

3 comments

Q1: Ablation on instance-centric formulation itself

Thank you for highlighting the need to isolate the contribution of the instance-centric formulation.

Table 5 (before revision) already includes a related ablation, "Predicting Empty-Space Scores." This variant predicts an additional score for empty space instead of fixing it to zero. As is discussed in Supplement D.1, because empty space is the complement of scene occupancy, predicting this score retains an explicit scene-level prediction task: the network can determine whether a point is occupied by adjusting the empty-space score, without relying entirely on the predicted instance geometries. This provides an alternative to learning occupancy through instance-level shape priors. Under the same supervision, this scene-centric variant achieves lower average RayIoU than ICOcc (10.9% versus 12.1%), supporting the benefit of deriving scene occupancy from instance-level predictions.

To provide a more direct comparison with conventional scene-centric prediction, we additionally evaluate a baseline that predicts scene density σ\sigma and computes opacity as α=1−exp⁡(−σδ)\alpha=1-\exp(-\sigma\delta). A separate slot-assignment head predicts π=softmax⁡(s)\boldsymbol{\pi}=\operatorname{softmax}(\boldsymbol{s}), yielding slot probabilities pq=απqp_q=\alpha\pi_q. Thus, scene occupancy is predicted independently of its partition into regions. This baseline uses the same view groups, pseudo segmentation, pseudo depth, surface and free-space objectives as ICOcc.

ConfigurationOAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrow
Predicting Empty-Space Scores0.8020.6350.573
Scene density + slot-assignment head0.8030.6270.491
ICOcc0.8130.6720.606

We have added the new ablation to the revised manuscript.

Q2: Evidence about instance-level shape priors learning

Thank you for requesting more direct evidence that the queries learn instance-level 3D shape priors.

To examine this directly, we have added cross-scene visualizations of 3D occupancy predictions from fixed query IDs to the revised manuscript. Please refer to Fig.3 (d, e). Each row shows predictions from the same query across scenes, alongside the input images. In the visualizations, the same query predicts similarly structured vehicle geometries or road regions in different scenes, suggesting that queries capture geometric regularities shared across scenes.

The analysis already provided in Supplement Table 4 (before revision) also offers complementary quantitative evidence. We measure each query’s semantic composition across test frames and compare it with independently randomized query identities in each frame. This permutation preserves the predicted regions and their geometry but destroys cross-scene identity associations. The original queries achieve higher consistency (74.4% versus 38.0%) and a lower switching gap (19.9% versus 54.2%). These results show that query identities exhibit non-random cross-scene organization, complementing the geometric evidence in the added visualizations.

I appreciate the authors for the response. The above explanation makes sense to me. Therefore, I tend to maintain my score.

Thank you for the follow-up and for considering our clarifications and additional experiments.

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

This paper presents ICOcc, an instance-centric framework for unsupervised monocular 3D occupancy prediction. Instead of predicting a single scene-level density field, ICOcc represents each 3D point using probabilities over multiple occupancy slots and an empty-space state. A query-based Transformer decoder assigns point features to 32 learnable slots, while projection-guided feature modulation incorporates image location and depth information. During training, UnSAM+ masks and Monodepth2 predictions provide pseudo instance and geometric supervision across multiple views. Dataset-provided 3D bounding boxes are additionally used to retrieve side views containing objects that are also visible in the input view. Experiments on KITTI-360 and a custom CarlaOcc split show improvements in occupancy precision, invisible-empty recall, and selected class-wise RayIoU.

Critical strengths

S1. Intuitive Instance-Centric Motivation

  • Separating the scene into competing occupancy slots is an intuitive way to reduce the entanglement of different objects along the same camera ray. The connection among categorical occupancy, query-based decoding, and multi-view mask supervision is easy to understand.

S2. Consistent Improvements Within the Submitted Evaluation

  • Under the paper’s reproduced protocol, ICOcc improves occupancy precision and invisible-empty recall on both KITTI-360 and CarlaOcc. The gains are particularly large on the custom CarlaOcc split.
Critical weaknesses

W1. Insufficient SOTA Claim and Insufficiently Described Baseline Reproduction

  • The state-of-the-art claim relies on reproduced baseline results, but the paper only states that the methods were reproduced under “unified configurations.” It does not explain why the originally reported KITTI-360 results were not used, whether the baselines were retrained or only re-evaluated, or whether official code and checkpoints were used.

  • This issue is particularly important for ReBench. Its original paper reports OAcc/OPrec/IERec of 0.883/0.763/0.676, whereas this paper reports substantially lower values of 0.798/0.591/0.408. If the original results were directly comparable, ICOcc’s 0.813/0.672/0.606 would not outperform ReBench on any of the three metrics. Differences in spatial range, frustum and visibility masks, occupancy representation, sampling procedure, training configuration, or checkpoint may explain the discrepancy, but none is documented.

  • Re-evaluating SceneDINO is understandable because its original results primarily use semantic mIoU rather than the OAcc/OPrec/IERec and RayIoU metrics adopted here. However, SceneDINO provides public code, checkpoints, and evaluation configurations, and the paper does not state whether these official resources were used. Similarly, reproducing the methods on CarlaOcc is necessary because original results are unavailable, but this does not explain the replacement of the reported KITTI-360 results.

  • The authors should report the original and reproduced results side by side and specify the codebase, checkpoint, retraining procedure, dataset split, etc. Without this information, it is unclear whether the reported improvements reflect the proposed method or degraded baseline reproduction, and the SOTA claim is not adequately established.

W2. Annotation Dependency Weakens the “Unsupervised” Claim

  • Despite being presented as an unsupervised method, the full model explicitly uses dataset-provided 3D bounding boxes to identify the same instances across views and select side-camera images. These boxes are manually provided 3D instance annotations. Therefore, the main method is not fully unsupervised or annotation-free, even though it does not use dense occupancy labels. Simply avoiding voxel-level occupancy labels does not make the entire training pipeline unsupervised.

  • The box-free ablation shows that an annotation-free version is possible, but the main results and state-of-the-art claims are based on the version using 3D boxes. The paper should describe its setting more precisely as “occupancy-label-free” or weakly supervised by 3D boxes. Alternatively, the box-free variant should be used for the primary comparison, or the box-based view selection should be replaced with an automatic image-only approach.

W3. Weak Generalization Evaluation

  • The real-world evaluation is limited to KITTI-360, while the CarlaOcc experiment uses a custom split containing only Town10HD sequences. There is no cross-dataset evaluation, cross-town experiment, or zero-shot transfer to SemanticKITTI, although Rebench provides such an evaluation. Consequently, the claim that the slots learn transferable category-level shape priors is not strongly supported.
Questions for authors

Q1. Baseline Reproduction and SOTA Claim

  • Why were the originally reported KITTI-360 results replaced with reproduced results? For each baseline, please clarify whether it was retrained or only re-evaluated and whether official code and checkpoints were used. In particular, what causes the large ReBench discrepancy between the original 0.883/0.763/0.676 and the reproduced 0.798/0.591/0.408 for OAcc/OPrec/IERec?

Q2. Use of 3D Bounding-Box Annotations

  • How can the main method be considered unsupervised when dataset-provided 3D bounding boxes are used to identify co-visible instances and select training views? Please clarify the required annotation level and revise the terminology accordingly.

  • Can the authors report the box-free variant as the primary comparison under the same experimental settings or replace the 3D-box-based retrieval with an automatic image-only method?

Q3. Generalization Experiment

  • Can the authors provide cross-dataset or zero-shot transfer experiments, such as KITTI-360 to SemanticKITTI, KITTI-360 to CarlaOcc, or training and testing across different CARLA towns? The current evaluation on KITTI-360 and a custom split of only Town10HD is insufficient to establish broad generalization.
6 comments

Q1: Baseline Reproduction and SOTA Claim

Thank you for highlighting the insufficient description of our baseline evaluation. The ReBench results cited in your review and those in our paper use different evaluation ranges and are therefore not directly comparable. We extended the maximum evaluation distance from 20 m to 50 m to evaluate occupancy prediction beyond the near-field region. The 3-50 m range was specified in Supplement E.1, Detailed Evaluation Protocols, but its difference from the original protocol was insufficiently emphasized. We have clarified this distinction in the revised manuscript.

To enable direct comparison with previously reported results, we additionally evaluate ICOcc over the original 3-20 m range:

MethodOAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrow
BTS0.8700.7330.658
KDBTS0.8710.7460.682
KYN0.8610.7460.707
ViPOcc0.8750.7480.668
SceneDINO0.8570.7300.542
ReBench0.8830.7630.676
ICOcc0.8840.8080.755

ICOcc also outperforms the compared baselines under the original evaluation range, supporting our state-of-the-art claim.

Regarding checkpoint sources, we used official checkpoints for BTS, KYN, and ViPOcc, whose shared training sampling settings are largely consistent. For KDBTS and SceneDINO, we aligned their training sampler configurations with these settings and retrained them while retaining their method-specific designs. ReBench was retrained following its published training procedure because official pretrained weights were unavailable. We have added these checkpoint and training details to the revised manuscript.

All methods were re-evaluated using a common evaluator implemented following the Rebench protocol, with checkpoint loading adapted to accommodate each baseline. Apart from the evaluation-range extension, the frustum and visibility masks, opacity-based occupancy representation, and evaluation sampling procedure follow the same protocol across methods. These choices are detailed in Supplement E.1, and we have made their consistency explicit in the revision.

Q2: Use of 3D Bounding-Box Annotations

Thank you for pointing out the ambiguity in our previous terminology. In the revised manuscript, we have removed "unsupervised" when describing ICOcc and instead state that it learns without ground-truth 3D occupancy annotations.

Following your suggestion, we additionally provide a box-free variant to examine whether instance-anchored view groups can be constructed without any annotations. The preprocessing approach is replaced by a VFM-based alternative. SAM 2.1[1] generates category-agnostic regions, and RoMa v2[2] provides cross-view correspondences, which are verified using camera geometry to associate regions and select side views. This procedure uses only RGB images and known camera calibration/poses, without dataset-provided instance boxes, identities, or category labels.

[1]: SAM 2: Segment Anything in Images and Videos, ICLR 2025.

[2]: RoMa v2: Harder Better Faster Denser Feature Matching, ECCV 2026

View-selection strategyDataset instance annotationsOAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrowAvg. RayIoU (%) ↑\uparrow
Without instance-anchored retrievalNo0.8080.6570.5678.3
VFM-based retrievalNo0.8100.6680.5879.1
Box-based retrievalYes0.8130.6720.60612.1

VFM-based retrieval improves all metrics over the variant without retrieval, showing that view-group construction can benefit occupancy learning without dataset-provided instance annotations. Compared with box-based retrieval, it yields smaller differences in geometric occupancy metrics and a gap in average RayIoU. Thus, instance annotations are not necessary for obtaining retrieval gains, although they further improve performance, particularly in class-wise occupancy prediction. The box-free VFM-based retrieval results, noted as ICOcc(AF), are moved into the main comparison table alongside the full version, making the results under both annotation settings directly accessible.

We would also like to go further in clarifying the role of the boxes. The bounding-box annotations are used only during dataset preprocessing, where only geometric information from the metadata of selected static instances serves as anchors for view-group construction. Their instance ids do not supervise occupancy or instance predictions. The box-free ablation retains instance occupancy prediction after removing all dependence on dataset-provided boxes, showing that this capability does not require box-based retrieval.

Importantly, access to bounding boxes does not resolve the problem motivating our work: learning instance geometry without dense occupancy ground truth. Recovering and separating instance geometry from indirect image-based supervision therefore remains the central challenge addressed by our formulation and network.

Q3: Generalization Experiment

Thank you for highlighting the limited generalization evaluation. We have added zero-shot cross-dataset evaluations on CarlaOcc and SemanticKITTI using the model trained on KITTI-360, without fine-tuning. The results are reported below.

MethodCarlaOcc OAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrowSemanticKITTI OAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrow
BTS0.7080.1910.3530.7320.2740.417
KDBTS0.6760.1620.3570.7090.2530.390
KYN0.6900.1570.3450.7510.2930.468
ViPOcc0.5190.0860.1990.5580.1620.276
SceneDINO0.7700.2430.2850.7430.3050.392
ReBench0.4850.0800.1930.6220.1920.317
ICOcc0.8450.3220.5470.7940.2920.569

Although cross-dataset generalization is not the primary focus of this work, ICOcc also shows improvements in these zero-shot evaluations. It achieves the highest OAcc, OPrec, and IERec on CarlaOcc, and the highest OAcc and IERec on SemanticKITTI. These results suggest that the instance-centric design may help transfer learned object geometry across datasets. We have added these results to the revised manuscript.

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

This paper proposes ICOcc, an instance-centric framework for unsupervised monocular 3D occupancy prediction. It combines a categorical occupancy formulation, a query-based instance decoupling network, and instance-anchored multi-view supervision. The method reports improvements in occupancy precision and invisible-empty recall on KITTI-360 and CarlaOcc.

Critical strengths

1.The query-based architecture is reasonably motivated, and the reported improvements in occupancy precision and invisible-empty recall are promising. 2.The ablations demonstrate the practical effectiveness of several components, particularly instance-related supervision and explicit free-space constraints.

Critical weaknesses

1.The instance-anchored view selection explicitly relies on dataset-provided 3D bounding boxes, instance visibility relationships, and semantic category filtering (Sec. 3.4 and Appendix D.3). Although dense 3D occupancy annotations are not used, this introduces annotated 3D geometric and semantic information into the training pipeline. It therefore constitutes a materially different supervision setting from methods trained exclusively with multi-view images and camera geometry. Importantly, removing the annotation-guided view retrieval reduces the reported average RayIoU from 12.1% to 8.3%, demonstrating that this additional information contributes to the performance gains. The comparison with strictly annotation-free methods is therefore insufficiently controlled. The paper should explicitly characterize its supervision requirements and provide matched-supervision comparisons. 2.Missing computational efficiency analysis. 3.The derivation of the sampling-interval correction also deserves clarification. Eq. (4) yields a rational opacity function rather than the exponential opacity used in conventional volume rendering. The implications for sampling invariance and occupancy calibration are not sufficiently established.

Questions for authors

1.The instance-anchored view selection explicitly relies on dataset-provided 3D bounding boxes, instance visibility relationships, and semantic category filtering (Sec. 3.4 and Appendix D.3). Although dense 3D occupancy annotations are not used, this introduces annotated 3D geometric and semantic information into the training pipeline. It therefore constitutes a materially different supervision setting from methods trained exclusively with multi-view images and camera geometry. Importantly, removing the annotation-guided view retrieval reduces the reported average RayIoU from 12.1% to 8.3%, demonstrating that this additional information contributes to the performance gains. The comparison with strictly annotation-free methods is therefore insufficiently controlled. The paper should explicitly characterize its supervision requirements and provide matched-supervision comparisons. 2.Missing computational efficiency analysis. 3.The derivation of the sampling-interval correction also deserves clarification. Eq. (4) yields a rational opacity function rather than the exponential opacity used in conventional volume rendering. The implications for sampling invariance and occupancy calibration are not sufficiently established.

5 comments

Q1: Supervision requirements and controlled comparisons

Thank you for highlighting the need to clarify annotation requirements and control the comparisons. We agree that this additional information must be explicitly documented and controlled for when comparing against methods that do not use it.

Our existing ablation in Table 3 (before revision) removes side-view retrieval while retaining the remaining ICOcc components. To additionally evaluate the baseline under the same retrieval condition, we compare SceneDINO and ICOcc both with and without annotation-guided retrieval:

SettingMethodAnnotation-guided retrievalOAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrow
ASceneDINONo0.7930.6220.465
BSceneDINOYes0.7940.6200.436
CICOccNo0.8080.6570.567
DICOccYes0.8130.6720.606

ICOcc outperforms SceneDINO under both conditions (C versus A and D versus B). Removing retrieval (C) eliminates ICOcc's dependence on dataset-provided instance annotations while retaining higher OPrec\mathrm{O}_{\mathrm{Prec}} and IERec\mathrm{IE}_{\mathrm{Rec}} than all compared prior methods on KITTI-360. Retrieval improves ICOcc (D versus C) but does not consistently improve SceneDINO (B versus A), indicating that access to the selected views alone does not explain ICOcc's advantage.

For more comparisons, we additionally provide a box-free variant to examine whether instance-anchored view groups can be constructed without any annotations. Specifically, the preprocessing approach is replaced by a VFM-based alternative. SAM 2.1[1] generates category-agnostic regions, and RoMa v2[2] provides cross-view correspondences, which are verified using camera geometry to associate regions and select side views. This procedure uses only RGB images and known camera calibration/poses, without dataset-provided instance boxes, identities, or category labels.

[1]: SAM 2: Segment Anything in Images and Videos, ICLR 2025.

[2]: RoMa v2: Harder Better Faster Denser Feature Matching, ECCV 2026

View-selection strategyDataset instance annotationsOAcc ↑\uparrowOPrec ↑\uparrowIERec ↑\uparrowAvg. RayIoU (%) ↑\uparrow
Without instance-anchored retrievalNo0.8080.6570.5678.3
VFM-based retrievalNo0.8100.6680.5879.1
Box-based retrievalYes0.8130.6720.60612.1

VFM-based retrieval improves all metrics over the variant without retrieval, showing that view-group construction can benefit occupancy learning without dataset-provided instance annotations. Compared with box-based retrieval, it yields smaller differences in geometric occupancy metrics and a gap in average RayIoU. Thus, instance annotations are not necessary for obtaining retrieval gains, although they further improve performance, particularly in class-wise occupancy prediction.

We have added these results and analysis to the revised manuscript.

Q2: Computational efficiency analysis.

Following your suggestion, we provide an analysis of inference efficiency below. The experimental results are measured on a single RTX 4090 GPU.

methodsTotal params(M)Trainable params(M)FPS ↑\uparrow
BTS34.8934.8914.15
ViPOcc58.7658.7614.10
KDBTS34.8934.8914.09
KYN442.6835.391.23
SceneDINO100.3413.665.20
ICOcc(Ours)120.3033.712.48

Total parameters include all model parameters, whereas trainable parameters exclude the frozen pretrained components and count only those updated during training. ICOcc has 120.30M total parameters and 33.71M trainable parameters, achieving 2.48 FPS during inference. Its additional inference cost comes from the Instance Decoupling Network in Sec. 3.3, including projection-guided feature modulation and the query-based Transformer decoder. Although inference is slower than some baselines, we consider this cost acceptable for our primary objective of instance-level 3D reconstruction. We have added these results to the revised manuscript.

Q3: Clarification on the derivation of the sampling-interval correction

Thank you for pointing out these omissions.

Eq. (4), also presented as Eq. (14) in the supplement, is based on the categorical softmax parameterization in Eq. (12) (before revision). This parameterization jointly represents mutually exclusive occupancy slots and empty space. The rational opacity is a consequence of this choice.

The interval correction log⁡(δi)\log(\delta_i) follows the reasoning of Alpha Invariance[3] and Rebench[4], which discuss the dependence of learned density magnitudes on spatial scale and sampling intervals. Specifically, photometric supervision constrains density through the opacity used in rendering, allowing density and interval length to scale inversely while preserving opacity. Our formulation likewise accounts for this inverse-scaling property, as shown below. Let λi=∑qexp⁡(ℓq(xi))\lambda_i=\sum_q\exp(\ell_q(\boldsymbol{x}_i)). Our opacity is αi=δiλi1+δiλi.\alpha_i=\frac{\delta_i\lambda_i}{1+\delta_i\lambda_i}. Under inverse scaling, δi′=kδi\delta_i'=k\delta_i and λi′=λi/k\lambda_i'=\lambda_i/k, we obtain

=\frac{(k\delta_i)(\lambda_i/k)} {1+(k\delta_i)(\lambda_i/k)} =\alpha_i.$$ Thus, our parameterization preserves the sampling invariance. [3]: Alpha Invariance: On Inverse Scaling between Distance and Volume Density in Neural Radiance Fields, CVPR24 [4]: Rebenchmarking Unsupervised Monocular 3D Occupancy Prediction, arXiv preprint arXiv:2602.06488 For occupancy calibration, adding $\log\delta_i$ converts scale-dependent scores into interval opacities through $\exp(\ell_q+\log\delta_i)=\delta_i\exp(\ell_q)$, as is discussed in Rebench. Under the inverse scaling described above, both the opacities and the resulting occupancy decisions at a fixed threshold remain unchanged. Following Rebench, we map these opacities onto voxel centers and apply a threshold of 0.5, using the same interval correction during training and evaluation. It is worth noting that the rational and exponential forms are first-order consistent. Writing $z_i=\lambda_i\delta_i$, $$\alpha_i^{\mathrm{rat}} =\frac{z_i}{1+z_i} =z_i-z_i^2+O(z_i^3), \qquad \alpha_i^{\mathrm{exp}} =1-e^{-z_i} =z_i-\frac12z_i^2+O(z_i^3).$$ We have incorporated these clarifications into the relevant sections of the revised manuscript to address the omissions the reviewer identified.

As the discussion deadline approaches, we have addressed your concerns regarding annotation requirements and controlled comparisons, inference efficiency, and the sampling-interval correction through additional experiments and clarifications in our response and revised manuscript. We would appreciate any remaining questions and would be happy to clarify further. If these additions address your concerns, we would be grateful if you could reconsider your assessment. Thank you again for your time and constructive feedback.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. In the revision, we clarified ICOcc’s annotation requirements and removed "unsupervised" when describing our method. We added matched-retrieval comparisons and a VFM-based retrieval variant that requires no dataset-provided instance annotations. We also documented baseline reproduction and the evaluation-range change, and provided results under the original 3-20 m protocol, where ICOcc leads on all three reported metrics.

To isolate the method’s contribution, we added a scene-centric baseline with the same supervision, which performs below ICOcc. Cross-scene query visualizations and zero-shot evaluations on CarlaOcc and SemanticKITTI provide further evidence of learned geometric structure and transfer performance. We also reported inference efficiency and clarified the sampling-interval correction. Together, these revisions strengthen the experimental support while explicitly documenting annotation requirements and computational costs.

Reviewer jQEQ has confirmed that our explanation makes sense and maintained their score.

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

Related papers

Discussion (0)

Sign in to comment.