- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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.
- 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.
- 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
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:
| Setting | SCOPE-R50 | SW-TC-Small |
|---|---|---|
| Backbone/pretraining | ResNet-50, same Cascade Mask R-CNN COCO+NuImages initialization | Same |
| Frozen stages | Stem + first stage | Same |
| Image size | ||
| Temporal input | 4 historical + current (5 sets) | 7 historical + current (8 sets) |
| Training | 40 epochs; global batch 16 | 70 epochs; global batch 8 |
| Local evaluation | Same 4,219 samples, target-frame ego/grid/camera mask | Same |
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 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 +SPLF full SCOPE reduces average 0.31 0.25 0.20 m and average collision rate 0.32% 0.26% 0.23% (Table 2). Our evaluator uses prefix-averaged open-loop metrics with a fixed-orientation 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.
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.
† 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.