acceptodds
Under review as a conference paper at ICLR 2027

RoadOcc Learns When to Persist, Transport, or Refresh Memory for Roadside Occupancy Prediction

Abstract

Fixed roadside cameras repeatedly observe a stable scene overlaid by sparse moving traffic. Temporal memory can recover weak observations, but reusing moving evidence at stale locations can corrupt occupancy predictions. Motion compensation addresses displacement, while reliance on the resulting history remains a separate learning problem. We introduce RoadOcc, which learns soft routing among fixed-coordinate history (Persist), velocity-addressed history (Transport), and current evidence (Refresh). Motion state and class-consistent historical support supervise these source choices. Dynamic-aware cross-attention (DCA) updates candidate locations, multi-scale voxel velocity estimation (VVE) constructs transport addresses from multi-scale current–history correspondence, and velocity-guided dynamic sparse fusion (VDSF) combines routed evidence under fixed sparse-token budgets. On InfraOcc, RoadOcc reaches 65.29 mIoU and 32.37 dynamic mIoU, gains of 4.44 and 4.71 over STCOcc. Controlled address experiments show that VVE raises dynamic mIoU by 0.87 over fixed-coordinate reading. Across three seeds, supervised P/T/R adds 1.40 dynamic points over motion-corrected retrieval, while removing Refresh costs 0.32 points. Results from two transfer models, Occ3D-nuScenes, and longer intervals provide additional support. Code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
roadside perception

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 4 reviews, above 68% of the others.
Reviewer_V91VOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RoadOcc predicts occupancy from fixed roadside cameras using current and historical features. Its routing module combines fixed-coordinate history (Persist), motion-compensated history (Transport), and current observations (Refresh), supervised by motion state and semantic support. It achieves 65.29 mIoU and 32.37 dynamic mIoU on InfraOcc, outperforming STCOcc.

Critical strengths
  • I find the P/T/R formulation well motivated for fixed roadside cameras. It separates motion compensation from the decision to reuse history.
  • The evaluation is reasonably extensive, including component and routing ablations, transfer to other temporal models, and tests on Occ3D-nuScenes and with longer frame intervals.
  • The paper is clearly written and easy to follow. The visualizations help explain motion compensation and the behavior of the three routes.
Critical weaknesses
  1. Table 6 shows that supervised routing helps, but the contribution of semantic support beyond motion-state supervision is harder to assess. A motion-only labeling baseline would be useful, particularly for slow objects whose P/T candidates may have similar support. Sensitivity to the support radius and 0.8 m/s velocity threshold is not reported. The paper acknowledges that support does not distinguish same-class instances, but does not evaluate the effect on history reuse.
  2. In Table 2, Other improves by 18.82 IoU points, accounting for roughly 30% of the overall mIoU gain. It is unclear which components produce this large improvement on a static class. Excluding Other still leaves a 3.34-point gain over STCOcc. Per-class voxel counts and results before and after adding P/T/R would make this result easier to interpret.
  3. The baseline adaptations need more detail. CRT-Fusion was developed for radar-camera detection, and Let Occ Flow uses self-supervised training. Please clarify the architecture and supervision used to adapt them. ST-Occ and GDFusion are cited but not evaluated. The Occ3D-nuScenes and temporal-gap tests assess the full model, leaving the generalization of P/T/R itself less clear.
Questions for authors
  1. For slow objects, how often do P and T yield different support outcomes? How sensitive are the results to the support radius and velocity threshold?
  2. Which components explain the large gain on Other? Can you provide per-class results before and after adding P/T/R?
  3. How were CRT-Fusion and Let Occ Flow adapted for camera-only occupancy prediction in terms of architecture and training supervision?
Reviewer_KqbbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RoadOcc, a new method for camera-based roadside 3D occupancy prediction. It studies camera-only 3D semantic occupancy prediction from fixed roadside cameras. The starting observation is that a roadside camera does not move, so history already lies on the current voxel grid. Static structure can therefore be reused as-is, but a moving object leaves a stale footprint at its old position. The paper turns this into a per-voxel choice among evidence of Persist, Transport, Refresh. A soft router mixes the three sources. It is trained with a rule-based target built from ground-truth motion and from whether the nearest previous label has the same class near the tested address.

Critical strengths

The paper measures the "static/dynamic asymmetry" problem related to speed, which clearly motivates the problem formulation. The core idea is to use a static roadside camera. This camera-only InfraOcc result is novel.

Critical weaknesses

The paper's central narrative emphasises dynamic evidence routing (Persist/Transport/Refresh), yet much of the reported performance improvement appears to come from factors outside the routing mechanism itself and often manifests as static occupancy gains. Table 4 shows that the proposed P/T/R routing contributes only +1.12 mIoU of the total +4.44 gain over STCOcc, with DCA and VVE accounting for most of it. Furthermore, several comparisons (e.g., CRT-Fusion + VDSF/PTR) show larger improvements in static than dynamic mIoU. This raises the question of whether the method's practical benefit comes primarily from dynamic routing or from the surrounding architectural changes.

The routing target is close to a speed threshold, and neither a rule-based router nor the Refresh route is evaluated directly. This raises questions about the value of the routing-target design.

The novelty seems mainly on the route target. VVE's BEV cost volume is close to ALOcc's cost-volume flow prediction (arXiv:2411.07725), and STCOcc already predicts scene flow jointly with occupancy (arXiv:2504.19749). CRT-Fusion warps past BEV features along predicted motion, and AsyncBEV warps BEV features along estimated flow (arXiv:2601.12994). GDFusion calibrates temporal fusion with a motion cue, and ST-Occ uses dynamic-aware memory attention (arXiv:2508.04705). CVT-Occ (ECCV 2024; arXiv:2409.13430) builds a temporal cost volume for occupancy and is not cited.

Unexplained sampling choice. The DCA threshold is drawn from a normal distribution with mean 0.5 and standard deviation 1, truncated to [0, 1] (lines 223 224). That is close to uniform on [0, 1]. Is that the intent?

Hard first read. The main text depends on many acronyms (DCA, VVE, VDSF, P/T/R, DSR, TP-MAVE, Nxt.) and on frequent appendix references. Tables 5 and6 and the text use "dMAVE" and "dM", but Section 4 defines only "Direct MAVE".

Questions for authors

Report static mIoU for every row in Tables 4 and 6. Is the router applied to static voxels in the top-K set, and what do its routes look like there? What moves the "Other" class from 68.84 to 87.66 (in Table 2)?

Which Table 2 rows are copied from the InfraOcc paper, and which did you train, with which backbone, input size, history length and schedule? Is "STCOcc-S" that paper's "Plain (C)" row, and what does "-S" denote for CRT-Fusion-S? What makes your STCOcc 11.15 mIoU stronger than the benchmark paper's (60.85 vs 49.70)? Which CRT-Fusion code did you use, and how did you adapt it to camera-only occupancy?

Please reconcile Table 5, Figure A7 and Table A7. What does each "read" variant change, which rows include P/T/R routing, and what is the "target-vector trace" configuration?

In OCC3D-NUSCENES, how are Persist and the P/T/R target defined when the ego vehicle moves, and which flow labels supervise VVE there?

Reviewer_bDF4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

RoadOcc predicts temporal semantic occupancy from four fixed roadside cameras. The scene is a stable scaffold with sparse short-lived traffic, so history helps weak or occluded actors but can corrupt predictions when it is reused at stale locations. The paper turns this into a per-voxel choice among three evidence sources: Persist, the history at the fixed coordinate; Transport, the history read at the velocity-addressed location; and Refresh, the current observation alone. Dynamic-aware cross-attention, DCA, uses the disagreement between the current and static predictions plus the dynamic probability to decide where image evidence is refreshed. Multi-scale voxel velocity estimation, VVE, builds the Transport address from a local cost volume between current and historical BEV slots and refines it from coarse to fine. Velocity-guided dynamic sparse fusion, VDSF, selects top-K tokens and mixes the three sources under a fixed budget. The route target comes from ground-truth motion and same-class support within a one-voxel radius, and the router is trained with equal-class-weight cross-entropy plus a soft-overlap term.

Critical strengths
  1. The fixed-roadside setting is framed correctly, and the Persist-Transport-Refresh split makes "where history should be read" and "how much to trust it" separately measurable. The fixed-coordinate and VVE rows of Table 5 differ only in the read address, and Table 6 varies only the routing variant, so the two factors are not confounded. The authors report seed-level standard deviations and note that the 0.021 m/s motion change is close to the 0.019 and 0.020 m/s seed standard deviations, which lets a reader judge the result.

  2. The evaluation contract is more complete than comparable papers. Dynamic and next-frame IoU, static and geometric guardrails, RayIoU, Direct MAVE, TP-MAVE, and DSR are reported together, and the paper explains why TP-MAVE must be read with recall. The token-budget saturation study and the temporal-gap test show that the result is not driven by an ever-growing budget and does not collapse when the temporal gap changes.

Critical weaknesses
  1. The paper reads as a module-by-module engineering pipeline, and it never shows that the learned router adds anything beyond the rule it imitates. The title promises a model that learns when to persist, transport, or refresh, but Equation 7 builds the route target from ground-truth motion and same-class support by a deterministic rule, so the network imitates a hand-written policy. The comparison that follows from this is missing. At inference the model already predicts velocity and occupancy, so the same rule can be evaluated on those predictions and hard-select Persist, Transport, or Refresh without the router network. If that variant matched the full model, the learned router would add little; if it failed, the value of learning would be visible. The paper reports neither this variant nor a simple motion-warped fusion baseline outside the DCA-VVE-VDSF stack. The closest control, the VVE read in Table 5, reaches 30.97 dynamic mIoU against 32.37, but it still uses the learned network around the address instead of the rule.

  2. The ablation never shows what routing does to the output. P/T/R is the claimed main innovation, yet the routing figures in Figures 7 and A5 are route maps conditioned on ground-truth dynamic support, and the paper itself says the maps separate source choices from occupancy detection. No figure compares the same scenes under no routing, gated mixing, and full routing, so a reader cannot see where the route decision changes the occupancy result, which object is recovered or damaged, or whether the three sources behave as claimed at object boundaries. A visual ablation on identical scenes would make the effect of routing visible instead of leaving it to aggregate tables.

Questions for authors
  1. The router imitates the rule in Equation 7, and at inference the model already predicts velocity and occupancy. Could you replace the router with the same rule evaluated on those predictions, hard-select one source per voxel, and report dynamic mIoU, next-frame IoU, Direct MAVE, and DSR? This would show whether the learned router is needed. A simple motion-warped fusion baseline outside your architecture would also show how much of the gain comes from the routing framework.

  2. Could you add a visual ablation on identical scenes that renders the occupancy output under no routing, gated mixing, and full routing, so the effect of P/T/R becomes visible at object boundaries and in the cases where the paper says routing matters?

Reviewer_yCBoOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes RoadOcc, a temporal occupancy prediction framework for roadside perception. Unlike ego-centric occupancy prediction, roadside cameras observe a largely stationary environment with sparse moving objects, making temporal feature reuse particularly important. To address this issue, the authors introduce three modules: Dynamic-Aware Cross-Attention (DCA), Voxel Velocity Estimation (VVE), and Velocity-guided Dynamic Sparse Fusion (VDSF). The key idea is to explicitly learn when to Persist, Transport, or Refresh historical features based on motion and semantic consistency. Experiments on InfraOcc demonstrate improvements over existing methods, achieving 65.29% mIoU and 32.37% dynamic mIoU.

Critical strengths
  1. Roadside occupancy prediction is a relatively new and underexplored setting. The paper addresses a practically relevant problem that differs from conventional ego-centric occupancy prediction.
  2. The distinction between persistent static structures and transient dynamic objects provides a reasonable motivation for selective temporal feature reuse.
  3. RoadOcc achieves strong performance on InfraOcc, particularly for dynamic occupancy prediction. The ablation studies also provide evidence supporting the proposed components.
Critical weaknesses
  1. Many methods in Table 2 were originally developed for different datasets or tasks rather than InfraOcc. However, the paper does not explain how these methods were adapted and evaluated on this benchmark. It is also unclear which results were reproduced by the authors and which were taken from existing benchmark evaluations. Several baseline numbers appear identical to those reported in the original InfraOcc paper. The authors should explicitly clarify the source and implementation of each baseline.
  2. Table 2 does not provide sufficient information about the configurations of the compared methods, including image backbones, input resolutions, camera views, temporal frames, and training schedules. In particular, RoadOcc uses different history lengths and sparse-token budgets from STCOcc. Without a detailed configuration comparison, it is difficult to assess whether the reported improvements are obtained under comparable settings.
  3. DCA, VVE, and VDSF largely build upon existing techniques, including sparse attention, motion compensation, and temporal feature fusion. The main novelty lies in the supervised P/T/R routing mechanism. However, this mechanism provides only a 1.40-point improvement in dynamic mIoU over the corresponding variant without routing. The larger improvement over STCOcc cannot be attributed solely to this contribution.
  4. The P/T/R routing targets are constructed using ground-truth occupancy and motion information derived from object annotations and temporal tracklets. It remains unclear whether the compared methods benefit from equivalent supervision. Additional experiments under matched supervision settings would help establish the contribution of the proposed routing mechanism.
Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.