acceptodds
Under review as a conference paper at ICLR 2027

SCOPE-Occ: Scene-Conditioned Probe Lifecycles for Trajectory-Free Camera-Only Occupancy World Modeling

33 views

Abstract

Camera-only occupancy world models forecast future 3D semantic scenes from historical observations. Existing approaches often condition on future ego trajectories, which require an upstream planner or predictor and may introduce oracle dependence or compounding errors. Without such inputs, models must infer how scenes evolve and when and where representational capacity is needed. We present SCOPE-Occ, a trajectory-free camera-only occupancy world model that treats sparse queries as scene-conditioned sensing probes. Its Scene-Conditioned Probe Lifecycle Field (SPLF) predicts temporal support functions governing each probe's emergence, persistence, and termination, supervised by probe-level support matching and scene-level capacity calibration. These signals route spatial evolution in Lifecycle-Routed Probe Dynamics (LPD), where persistent and emerging probes follow distinct update paths integrating observed ego-motion context, local visual evidence, learned mobility, and lifecycle-gated relational interactions. Existence scores further gate probe contributions to future occupancy decoding. On Occ3D-nuScenes, SCOPE-Occ outperforms prior trajectory-free camera-only methods, achieving 21.68 mIoU and 45.26 IoU averaged over the 1-3s forecasting horizons, without future ego trajectories at inference.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Occupancy Forecasting, World Modeling, Camera-only Perception, Scene-conditioned Representations, Lifecycle Modeling, Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_mZFeOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies camera-only semantic occupancy forecasting at 1–3 s without future ego trajectories or future poses. SCOPE keeps a shared set of sparse probes, predicts a scene-conditioned temporal support curve for each probe, and uses the resulting persistence and emergence signals to route spatial updates and gate voxelization. The central claim is that this organization improves long-horizon forecasting enough to outperform prior trajectory-free camera-only methods and to approach predicted-trajectory references. SCOPE achieves strong trajectory-free camera-only occupancy forecasting results on Occ3D-nuScenes.

Critical strengths
  1. The trajectory-free protocol is concrete: Section 4.1 restricts inference inputs to historical and current images, calibration, and historical ego states, and it states that future poses are used only offline to build labels.
  2. The mechanism is specified, not only named. Eq. 4 is a smooth support window; Eq. 5 splits it into inherited support p and newly activated support γ with p+γ=e; Eq. 12 sends those two signals down different displacement paths; Eq. 13 gates voxel scores by e.
  3. Table 9 shows why the problem is real for this architecture family: SparseWorld-TC-Small falls from 25.60 average mIoU with ground-truth future trajectories to 21.60 with predicted trajectories and 15.99 with the trajectory removed, and 0.25–1.0 m or 2–5° perturbations degrade it further. SCOPE at 21.02 is measured in that no-trajectory column.
Critical weaknesses
  1. The paper claims a world model, but the method and the experiments do not implement one. The title, abstract, and Section 2.2 place SCOPE in the occupancy world-model line, whose job is to predict how the scene changes over time. In that line, future ego trajectory is the action: SparseWorld-TC, DOME, and COME condition scene evolution on it, and OccWorld rolls scene tokens together with ego tokens. A world model here means a transition model, future scene given current scene and an action. SCOPE removes the action, which is more like a 4D occupancy forecaster. Section 3.1 and Section 4.1 pass in only historical images, calibration, and historical ego states; future trajectories and future poses never enter the network. Section 3.3 decodes all horizons in parallel, so a predicted motion is never applied and rolled forward.
  2. The forecasting claim is also unseparated from static occupancy. Section 3.1 puts every future target in the current reference frame, so static structure largely stays fixed and copying the current occupancy is a strong baseline. That baseline is absent. There is no per-class table and no dynamic-only mIoU. Occ3D mIoU and IoU are dominated by large static classes, so the gains in Tables 3–4 can come from covering drivable surface and vegetation rather than from predicting how objects move. Figure 3 compares methods on one scene; Figure 6 shows SCOPE alone. Without a dynamic split, the paper does not show that probe lifecycles model scene evolution.
  3. The headline comparison does not control the backbone. The abstract, introduction, and conclusion lead with 21.68 mIoU and 45.26 IoU, which Table 1 attributes to DINOv3 ConvNeXt-Base. The ResNet-50 model is 21.02 mIoU and 42.99 IoU. No trajectory-free DINOv3 baseline is reported. Appendix B.1 initializes the ResNet-50 encoder from a Cascade Mask R-CNN pretrained on COCO and NuImages, while the † reproductions are not stated to share that initialization, the 256×704 resolution, the 40-epoch schedule, or the history length. Rows without †, including SparseWorld at 13.20 mIoU, are not identified as re-runs. The +5.03 mIoU over trajectory-free SparseWorld-TC-Small† is large, and it is not a matched comparison.
  4. The lifecycle targets do not track a region across time. Section 3.2 and Appendix A.2 set y_i^h from an independent 1-NN assignment at each horizon. Equation 14 then treats the displacement between centroids of the same slot index as mobility. Appendix F.3 states that this assignment does not enforce semantic or instance consistency. The mobility term still adds 1.44 average mIoU in Table 3, so the extra pathway helps the metric, but the paper’s reading of m as motion of “the probe-supported region” is stronger than the supervision.
  5. Table 2 does not support the safety sentence in Section 4.2. The trajectory head does not condition occupancy, and the paper does not state that its L2 and collision protocol match UniAD or OccWorld. SCOPE’s 1 s collision rate is 0.24, above UniAD (0.05), OccWorld (0.12), and SparseWorld (0.11). Its 1 s L2 of 0.16 m is also far below UniAD’s 0.48 m, which indicates a mismatched open-loop protocol rather than a demonstrated planning result.
Questions for authors

I think the authors should reconsider the scope of their algorithm, which in my opinion belongs to a 4D occupancy forecaster rather than a world model in the usual sense, mainly because it does not model actions.

9 comments
AuthorsOpenReview ↗

Method SCOPE and Scene Evolution

We sincerely thank the reviewer for the detailed critique of the task scope, methodological evidence, and evaluation. We agree that several original claims were too broad. We have revised those claims and added the requested static-persistence, change-region, and multi-scene comparisons.

W1 — Occupancy forecasting is not an action-conditioned world model

We agree with this distinction. SCOPE is a trajectory-free camera-only occupancy forecaster, not a model of counterfactual scene transitions conditioned on proposed ego actions. We changed the title to SCOPE-Occ: Scene-Conditioned Probe Lifecycles for Trajectory-Free Camera-Only Occupancy Forecasting, and aligned the Abstract, Introduction, Sec. 3.1, and Conclusion. The method predicts target horizons in parallel from historical/current images, calibration, and ego states; its routed geometric updates do not constitute action-conditioned rollouts. This is an intentional scope limitation, not a claimed alternative to controllable planning world models.

W2 — Does the method predict changes beyond static occupancy?

We added an explicit oracle-motion static-persistence diagnostic: propagate Full SCOPE's predicted current occupancy, assuming an unchanged world, into each future ego frame using GT future ego poses. This reference has privileged information and is not claimed as a fair trajectory-free inference baseline. All models are compared over paired-valid voxels, where current/future GT are spatially aligned and both camera masks and grid overlap are valid.

Method (paired-valid)Overall mIoU ↑Overall IoU ↑Changed mIoU ↑New rec. (%) ↑Vac. rec. (%) ↑
Static persistence$^{\dagger}$24.4747.1812.5331.8547.16
SW-TC, no future input17.8938.1412.7628.3660.27
Base16.4336.7111.6744.6656.27
Base + SPLF19.7440.6417.3353.1964.45
Full SCOPE22.5644.2419.7955.3267.72

†^{\dagger} Uses GT future ego poses. We explicitly acknowledge that oracle persistence is better in overall mIoU (24.47 vs. 22.56), reflecting the strength of static scene preservation. Nevertheless, SCOPE obtains +7.26 changed-region mIoU, +23.47 pp newly occupied recall, and +20.56 pp vacated recall. It also exceeds no-trajectory SW-TC in changed-region mIoU (19.79 vs. 12.76). The 17-class breakdown shows SCOPE improves on Base in every class; persistence remains stronger on many structural categories. We added Table 5 and Appendix D.5/Table 14, plus four side-by-side multi-scene cases (Figs. 7–8), including a turning-bus failure. Their highlighted areas fall outside paired-valid evaluation regions, so we do not present them as direct visualizations of the diagnostic metrics. These results support forecasting of GT-defined semantic occupancy changes, not instance-level motion tracking.

AuthorsOpenReview ↗

Evaluation, Fairness, Mobility, and Planning

W3 — Backbone, protocol, and training fairness

We agree that the DINOv3 result should not be the primary evidence for a method-level improvement. Our headline methodological comparison instead uses SCOPE-R50 and locally evaluated SW-TC-Small-R50 under the same final scoring protocol. The following details are now explicit:

SettingSCOPE-R50SW-TC-Small
Backbone/pretrainingResNet-50, same Cascade Mask R-CNN COCO+NuImages initializationSame
Frozen stagesStem + first stageSame
Image size256×704256\times704256×704256\times704
Temporal input4 historical + current (5 sets)7 historical + current (8 sets)
Training40 epochs; global batch 1670 epochs; global batch 8
Local evaluationSame 4,219 samples, target-frame ego/grid/camera maskSame

Under the no-future-input condition, SCOPE-R50 improves average mIoU/IoU from 15.99/36.91 to 21.02/42.99. We do not claim identical training compute or a fully controlled external ablation. Our mechanism attribution instead uses the within-backbone component ablations (Tables 3–4), including the matched-supervision gate control; DINOv3 is presented only as a stronger optional backbone configuration. Original published results under nonidentical or unverified evaluation protocols have been moved to Appendix C.1 (Tables 7–8); training differences are documented in Appendix B.1.

W4 — Does a probe slot track a physical region?

We agree that shared probe indices provide architectural slot correspondence, not identity or region tracking. At each horizon, 1-NN GT support matches are constructed independently. Occupied GT voxel centers are converted into metric coordinates and aligned to the current ego frame before matching candidate points in that same frame. The displacement between matched-support centroids is therefore a weak spatial-support shift proxy; changes in the supported GT subset can affect it. It should not be interpreted as physical velocity or an instance-consistent motion target. The first pair h=0→1h=0\rightarrow1 also uses the same support-matching and reliability rules. We revised Sec. 3.3 (Eqs. 9, 14), Appendices A.1–A.4 and G.2. Any gain from the mobility route indicates predictive utility, not validation of object-level tracking.

W5 — Do auxiliary trajectory scores establish safe planning?

No. We removed the original claim of ``safe long-horizon trajectory prediction.'' The auxiliary head is trained only on detached features from frozen occupancy models and does not influence occupancy forecasts. Under an identical head/evaluator, Base →\rightarrow +SPLF →\rightarrow full SCOPE reduces average L2L_2 0.31 →\rightarrow 0.25 →\rightarrow 0.20 m and average collision rate 0.32% →\rightarrow 0.26% →\rightarrow 0.23% (Table 2). Our evaluator uses prefix-averaged open-loop metrics with a fixed-orientation 4.084×1.85 m4.084\times1.85\,\mathrm m ego box and GT vehicle/foreground and pedestrian masks. At GT-collision timestamps, predicted collision counts are suppressed in the numerator but timestamps remain in the denominator. Published UniAD/OccWorld/other scores use their reported protocols and are contextual references, not a controlled ranking (Sec. 4.1; Appendices B.5 and C.2). We make no closed-loop planning or safety claims.

The authors have shifted the topic of the paper from "world models" to "occupancy forecasting" and have implemented extensive revisions. In my view, "Weakness 1" has been addressed effectively. However, this change has introduced some new issues. The methods compared in Table 1 are all "trajectory-conditioned world models" rather than approaches designed for the "occupancy forecasting" task (the settings are different). The paper lacks comparisons with methods highly relevant to this task, such as Cam4DOcc and OccProphet, which are cited in the related work section. I suggest that the authors include additional comparative experiments involving these relevant methods.

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

This paper presents SCOPE-Occ, a camera-based method for forecasting future semantic occupancy without future ego trajectories or poses as inference inputs. Its main contribution is a scene-conditioned lifecycle model that determines the temporal participation of sparse probes. These lifecycle signals are subsequently used to route spatial refinement, mobility-aware relational aggregation, and occupancy decoding. Evaluation on Occ3D-nuScenes at 1, 2, and 3 seconds shows substantial improvements over the reported trajectory-free baselines, with performance approaching a predicted-trajectory-conditioned reference. Component ablations and additional analyses investigate the contributions of the proposed mechanisms.

Critical strengths
  1. The paper addresses a meaningful dependency in trajectory-conditioned occupancy forecasting. Ground-truth future ego trajectories are unavailable at deployment, while predicted trajectories introduce another source of error. SCOPE-Occ offers a coherent alternative that organizes sparse representations directly from historical observations. Its substantial gains over the reported trajectory-free baselines, including with the ResNet-50 backbone, provide meaningful empirical support for this design.

  2. The manuscript is clearly written and unusually explicit about the scope and limitations of its claims. In particular, it distinguishes probe support from object identity and acknowledges the limitations of weak matching, fixed computation, and discrete-horizon evaluation. The extensive ablations make the proposed components easier to assess and strengthen confidence in the reported empirical contribution.

Critical weaknesses
  1. The practical role of trajectory-free occupancy forecasting is insufficiently established. An unconditional scene forecast can be useful, but the paper does not clearly explain when its advantages outweigh the ability of trajectory-conditioned models to evaluate scene evolution under a proposed ego motion. The occupancy results establish forecasting accuracy under the chosen protocol, while the auxiliary trajectory-head results do not isolate whether the proposed occupancy representation improves driving decisions. This limits the demonstrated significance for autonomous driving.

  2. The distinct methodological contribution of lifecycle modeling is not fully isolated. The approach combines temporal confidence gating, auxiliary supervision, relational aggregation, and spatial refinement. Although the ablations show that these components are useful, they do not establish whether the lifecycle formulation is necessary compared with simpler scene-conditioned per-horizon gates under equivalent supervision and model capacity. For example, Table 4 does not include a horizon-wise gate with the same capacity-calibration loss. This leaves uncertainty about how much of the improvement depends on the proposed lifecycle structure.

Questions for authors
  1. Application and downstream evidence: What specific deployment setting motivates forecasting without future ego-trajectory conditioning? In Table 2, what features and ego-state inputs does the trajectory head receive, and is there evidence that the proposed future-occupancy mechanisms improve trajectory prediction compared with an otherwise identical model without them?

  2. Contribution isolation: How does a scene-conditioned per-horizon gate perform when trained with the same existence and capacity-calibration losses as SPLF, while keeping the remaining architecture and training settings fixed? This comparison would materially clarify the benefit of the lifecycle parameterization.

  3. Supervision clarification: Appendix A.2 states that each candidate point is independently assigned to an occupied ground-truth voxel through 1-NN matching. What rule makes a point unmatched when constructing the support target in Eq. (6)? Without a rejection rule, the stated assignment appears to give every candidate point a match, making the support target uniformly one. Please clarify the exact assignment procedure.

  4. Evaluation clarification: Are all results in Table 1 evaluated using the same current-reference coordinate frame, spatial region, and visibility-mask protocol, including results taken from prior papers? Please explain how protocol consistency is ensured.

These are requests for clarification and targeted evidence that could change my assessment, rather than suggestions for broader future work.

3 comments
AuthorsOpenReview ↗

Practical Relevance and Down Stream Evidence

We sincerely thank the reviewer for recognizing the motivation and for identifying where the original submission lacked controlled evidence. We have revised the paper and added the requested matched-supervision experiment, downstream ablations, and common-protocol comparisons.

Q1 — Deployment setting and downstream utility

Q1.1. When is trajectory-free forecasting useful? Our intended use is online predictive perception, where the latest scene forecast may be needed before an upstream planner or ego-trajectory predictor produces a future path. We agree that an unconditional forecast cannot replace an action-conditioned model for comparing alternative ego plans. Its distinct value is providing a timely forecast without that dependency. On a single RTX 4090, we measured:

ConfigurationMean latency (ms/sample) ↓
SW-TC-Small, no future trajectory105.94
SW-TC-Small, predicted trajectory including online trajectory generation255.10
SCOPE-R50, no future trajectory92.68

The measured SCOPE pipeline has 63.7% lower latency than the online-trajectory-conditioned SW-TC pipeline. These are pipeline-latency measurements, separate from the occupancy-only FPS in Table 1. We added the deployment setting to the Introduction and the latency comparison to Appendix E, Table 16.

Q1.2. What enters the auxiliary trajectory head? It predicts six future ego positions at 0.5-s intervals, using a 21-D observed ego-state vector: four historical pose encodings (12-D), velocity (3-D), acceleration (3-D), and a 3-D high-level navigation command (not future ground-truth motion). For each future step, the head receives the top-256 probe representations ranked by confidence ×\times existence. Each probe contributes 256-D features, 12-D point geometry, and 19-D semantics/support (17 class probabilities, confidence, existence). The head predicts a gated residual from an ego-state-based reference. Full details are in Appendix B.5.

Q1.3. Do the proposed occupancy features help downstream prediction? To isolate representation quality, we freeze each occupancy model, detach all probe inputs, and train the same auxiliary trajectory head with identical ego-state inputs, training, and evaluator. Its output is never fed back into occupancy forecasting.

Frozen representationL2L_2@1s@2s@3sAvg. L2L_2 (m) ↓Avg. collision (%) ↓
Base0.250.300.370.310.32
Base + SPLF0.200.250.310.250.26
Full SCOPE0.160.200.240.200.23

Thus, the proposed representations improve this auxiliary open-loop task under controlled conditions; we do not interpret these results as evidence of better closed-loop driving decisions. The ablation is now in Sec. 4.2, Table 2.

AuthorsOpenReview ↗

Lifecycle Isolation, Matching, and Evaluation

Q2 — Does the continuous lifecycle help beyond a horizon gate?

We added the requested scene-conditioned per-horizon gate, which predicts independent temporal support scores (including the current timestamp), and matched existence targets, capacity supervision when enabled, supervised horizons, LPD, and training settings. The two support heads have comparable, though not identical, sizes (0.30M gate vs. 0.33M lifecycle parameters).

Temporal supportCapacity lossAvg. mIoU ↑Avg. IoU ↑Capacity error ↓
Horizon gateNo18.2240.220.28
Continuous lifecycleNo19.8941.610.19
Horizon gateYes19.0140.970.17
Continuous lifecycleYes21.0242.990.13

Under matched capacity calibration, the continuous lifecycle gains +2.01 mIoU / +2.02 IoU over the gate; without it, gains remain +1.67 / +1.39. This provides direct evidence for the temporal-support parameterization, beyond capacity supervision alone. See Table 4 and Appendix A.2.

Q3 — What makes a 1-NN assignment unmatched?

We apologize for omitting the acceptance rule. For an occupied-GT center set Gh0\mathcal G_h^0, future GT voxel centers are aligned into the current ego frame; candidate points are decoded from normalized coordinates into the same metric frame. A candidate is accepted only if its nearest occupied GT center is within 0.8 m:

\|\bar{\mathbf p}_{i,n}^{h,0}-\mathbf g\|_2\le0.8\,\mathrm m\right], \qquad y_i^h=\frac1{64}\sum_{n=1}^{64}a_{i,n}^h.$$ When $\mathcal G_h^0=\emptyset$, all points are unmatched. A nearest neighbor beyond 0.8 m is likewise rejected; consequently, $y_i^h$ is not necessarily one. The separate $y_i^h\ge0.05$ threshold (at least 4/64 accepted points) selects probe-reliable slots **only for auxiliary dynamics losses**. The current timestamp $h=0$ uses the same match rule and contributes to the first adjacent pair. Clarified in **Sec. 3.2 and Appendices A.1–A.4**. ## Q4 — Are the occupancy results evaluated consistently? We agree the original Table 1 combined results that should not have been directly ranked. We corrected the coordinate description and locally reevaluated all baseline configurations now in Table 1 on the same 4,219 validation samples at each of 1/2/3 s, using the target timestamp's native ego frame, $200\times200\times16$ voxels at 0.4 m, and the target-frame Occ3D camera visibility mask. Methods are grouped by GT, predicted, or absent future ego-motion input at inference; local evaluation does not imply retraining each baseline. Under this protocol, SCOPE-R50: 21.02 mIoU / 42.99 IoU, versus SW-TC-Small without future input: 15.99 / 36.91; the latter with predicted future motion obtains 21.60 / 44.91. Published results with different/unverified protocols are presented separately (Appendix C.1, Tables 7–8). For example, the original SW-TC evaluation uses 5,719/5,419/5,119 samples at 1/2/3 s, not 4,219 at every horizon. We also disclose differences in historical inputs and training schedules (**Appendix B.1**) and do not claim a fully matched external training ablation.
AuthorsOpenReview ↗

Additional Results under Common Evaluation Protocols

Following our previous response to Q4, we have expanded the common-protocol comparison by adding PreWorld and OccProphet to Table 1. Both are locally evaluated on the same 4,219 Occ3D-nuScenes validation samples, using the same voxel grid, the target timestamp’s native ego frame, and target-frame camera visibility masks.

PreWorld uses its released occupancy-forecasting checkpoint. OccProphet is an adapted reproduction for the Occ3D semantic labels, voxel grid, and 1–3 s forecasting setup, as documented in Appendix C.1.

We have also added a cross-annotation evaluation on PanoBench-nuScenes in Appendix C.3 and Table 10. The results below are averaged over 1–3 s, without future ego-motion inputs:

MethodOcc3D Avg. mIoUOcc3D Avg. IoUPanoBench Avg. mIoUPanoBench Avg. IoU
PreWorld15.5238.9713.2627.42
OccProphet, adapted reproduction14.8037.6911.6928.02
SCOPE-R5021.0242.9915.6530.59

For PanoBench, all methods use frozen Occ3D-trained checkpoints without fine-tuning, and are evaluated on the same 5,119 prediction origins using the released target-frame valid masks. This provides an additional comparison under alternative annotations and visibility masks within the same nuScenes domain.

SCOPE-R50 outperforms both newly added baselines under each evaluation protocol. Appendix C.3 documents the PanoBench label mapping, grid alignment, and mask handling to make the comparison reproducible.

We would appreciate your feedback on whether these additional comparisons, together with our earlier responses, address your main concerns. As the discussion period draws to a close, we remain available to clarify any outstanding points.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Supervision Consistency and Scene-Change Anlysis

We appreciate the reviewer's careful inspection of the supervision code and request to identify the source of our gains. We have corrected the missing implementation details and added region-level and per-class analyses to the revision.

Q1 — Accepted matches, reliability, and effective training weights

Point matches versus probe reliability. The two thresholds serve different purposes. For each horizon, predicted candidate points and occupied GT voxel centers are expressed in the current metric ego frame. The detached 1-NN assignment accepts a candidate only when its nearest occupied GT center is within 0.8 m. The lifecycle target yihy_i^h is the fraction of accepted points among 64; when the eligible GT set is empty, its value is zero. Separately, yih≥0.05y_i^h\ge0.05 (at least 4/64 points) determines whether a slot is reliable enough for auxiliary mobility/emergence supervision. It does not redefine point matching or the lifecycle target. We now specify the rule in Eq. (6), Appendix A.2, and Appendix A.4.

λlife=0.8\lambda_{\mathrm{life}}=0.8 versus the supplied default 0.03. The reviewer is correct that the original supplementary base configuration contained scope_life_loss_weight=0.03. We regret that our submission omitted the effective experiment overrides: 0.03 was a legacy debugging default, not the setting of the reported checkpoints. We rechecked the checkpoints. The ResNet-50 and DINOv3 main experiments, and applicable component/lifecycle ablations in Tables 3–4, use λlife=0.8\lambda_{\mathrm{life}}=0.8 (weighting Lexist+Lcap\mathcal L_{\mathrm{exist}}+\mathcal L_{\mathrm{cap}} when both are enabled) and sm=2 ms_m=2\,\mathrm m. Disabling capacity loss removes Lcap\mathcal L_{\mathrm{cap}} while retaining the existence-loss weight. In the Table 12 sensitivity study, the λlife\lambda_{\mathrm{life}} sweep fixes sm=2 ms_m=2\,\mathrm m and the sms_m sweep fixes λlife=0.8\lambda_{\mathrm{life}}=0.8. These effective settings and the supplementary-code discrepancy are now explicitly documented in Appendix B.3.

Q2 — Are the improvements confined to static structure?

We added a paired-valid diagnostic using aligned current/future ground truth and the intersection of camera-valid masks and overlapping spatial grids. This avoids treating changes in spatial coverage as occupancy transitions. Semantic-change mIoU is measured on GT-defined semantic label changes, whereas newly occupied recall and vacated recall measure GT free$\rightarrow$occupied and occupied$\rightarrow$free transitions. Counts are accumulated across samples per horizon before averaging the 1/2/3-s scores.

Method (paired-valid)Overall mIoU ↑Changed mIoU ↑New recall (%) ↑Vacated recall (%) ↑
Oracle static persistence †^{\dagger}24.4712.5331.8547.16
SW-TC-Small, no future input17.8912.7628.3660.27
Base16.4311.6744.6656.27
Base + SPLF19.7417.3353.1964.45
Full SCOPE22.5619.7955.3267.72

†^{\dagger} The diagnostic propagates Full SCOPE's predicted current occupancy using GT future ego poses; it is not a trajectory-free inference baseline. On changed-region mIoU, the controlled progression is 11.67 →\rightarrow 17.33 →\rightarrow 19.79. Full SCOPE improves over Base in all 17 semantic classes under the same paired-valid protocol (Table 14); for example, car IoU rises from 15.25 to 23.36. We do not equate GT semantic changes with tracked-object motion. The paired-valid diagnosis is separate from the standard Table 1 protocol; Figs. 7–8 are complementary full-scene illustrations, not views of the evaluated paired-valid voxels. See Sec. 4.3, Table 5, Appendix D.5, Table 14, and Figs. 7–8.

AuthorsOpenReview ↗

Updated Manuscript with Revision Highlights

Dear Reviewers and Area Chair,

We have uploaded a highlighted version of the revised manuscript to make the changes discussed in our rebuttal easier to locate.

Blue text marks key revisions relative to the original submission. For readability, table entries and figure colors are not exhaustively recolored; changes involving tables and figures are explained in their captions, surrounding text, and our existing responses. The highlighting is intended as a reading aid rather than a word-by-word diff.

This upload adds revision highlighting and minor editorial corrections only. It introduces no additional experiments and does not change the reported results or conclusions from the revised manuscript accompanying our rebuttal. Our previously posted responses remain applicable.

For convenience, the main revisions discussed in those responses can be found at:

TopicLocation in the revised manuscript
Common-protocol occupancy comparisons and published-result provenanceTable 1; Appendix C.1, Tables 7–8
Matched-supervision lifecycle control and frozen-representation trajectory ablationsTables 4 and 2; Appendices A.2 and B.5
Static-persistence and scene-change diagnostics, per-class results, and qualitative/failure casesTables 5 and 14; Appendices D.5 and F.3; Figs. 7–8
Supervision definitions, effective training settings, and deployment latencySection 3; Appendix A; Appendix B.3; Appendix E, Table 16

Thank you for your time and consideration. We remain available to clarify any further questions.

AuthorsOpenReview ↗

Additional Component Analysis on PanoBench-nuScenes

To complement the change-region and per-class analyses in our response to Q2, we have added a cross-annotation evaluation on PanoBench-nuScenes (Appendix C.3, Table 10), which provides alternative annotations and visibility masks within the same nuScenes domain.

All configurations use frozen Occ3D-trained checkpoints, without fine-tuning or future ego-motion inputs. They are evaluated on the same 5,119 prediction origins using the released target-frame valid masks. The scene-wide results, averaged over 1–3 s, are:

ConfigurationAvg. mIoU ↑Avg. IoU ↑
Base11.1724.98
Base + SPLF13.7727.43
Full SCOPE15.6530.59

Adding SPLF improves average mIoU and IoU by 2.60 and 2.45 points, respectively; adding LPD provides further gains of 1.88 and 3.16 points. Both components improve the results at every forecasting horizon.

This new evaluation complements the GT-defined change-region analysis previously reported in Table 5: the earlier analysis identifies improvements on semantic occupancy changes, while the additional results show that the component gains also hold under a different annotation and visibility-mask protocol.

We would appreciate your feedback on whether this additional analysis and our earlier clarifications address your questions about the supervision and the source of the gains. We would be happy to clarify any remaining concerns before the discussion period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.