PaperlayerView exact revision on arXiv

Source-linked AI summary

PanoBench: A Panoptic Occupancy Prediction Benchmark for Autonomous Driving

Yi Feng, Zizhan Guo, Junwu E, Yu Ma, Mengtan Zhang, Jin Wu, Wei Ye, Hanli Wang, Qijun Chen, Danwei Wang, Rui Fan

arXiv:2610.23537v1cs.CV

TL;DR

Existing occupancy benchmarks often lack complete instance-aware annotations, while synthetic panoptic data alone does not capture real-world driving characteristics. PanoBench combines synthetic and reconstructed real-world data under a unified protocol, and its analyses find improved annotation quality alongside persistent challenges in real-world instance-aware prediction.

  • Problem

    Existing real-world occupancy annotations are incomplete and difficult to extend to reliable instance-level supervision, while synthetic-only benchmarks do not capture real-world driving diversity.

    Source: §I · p. 1

  • Method

    PanoBench unifies synthetic mesh-derived and reconstructed real-world panoptic occupancy datasets with common construction and evaluation protocols.

    Source: §I · pp. 1–2

  • Results

    PanoBench annotations are more observation-faithful, complete, and less fragmented than existing benchmarks, while no single architecture dominates across metrics and real-world instance-aware prediction remains challenging.

    Source: 4) Discussion · p. 15, §V · p. 16

  • Takeaways & Limitations

    Volumetric completeness and surface precision are distinct objectives, so IoU alone can mis-rank methods and combining dense coverage with sparse surface precision is promising.

    Source: 4) Discussion · p. 15

    Limited camera coverage, as in PanoBench-KITTI360, hampers current and future prediction.

    Source: 4) Discussion · p. 15

Abstract

from arXiv · show

Panoptic occupancy unifies 3D geometry, semantics, and instance identities, but its development is limited by scarce instance-aware annotations. Existing benchmarks mainly provide semantic labels reconstructed from incomplete sensor observations, while panoptic datasets have limited domain coverage or annotation completeness. We introduce PanoBench, a camera-centric panoptic occupancy prediction benchmark spanning synthetic and real-world driving scenes. It combines PanoBench-CarlaOcc, PanoBench-nuScenes, and PanoBench-KITTI360 through canonical data interfaces and unified evaluation protocols, providing multi-resolution annotations with persistent instance identities. The synthetic branch derives complete panoptic geometry from simulated meshes. For real-world data, a new pipeline reconstructs backgrounds using geometric priors and topology constraints, generates foreground mesh assets from multi-frame multi-view images, and fuses them into sparse panoptic voxels. A unified valid-mask protocol evaluates complete thing instances supported by temporal observations while scoring stuff regions at the voxel level. Analyses of observation fidelity, geometric completeness, spatial continuity, and temporal consistency demonstrate the benefits of the resulting annotations over existing occupancy benchmarks. We benchmark representative methods on semantic occupancy prediction, panoptic occupancy prediction, and 4D occupancy forecasting, and study voxel resolution, synthetic-to-real transfer, and storage efficiency. These results reveal complementary strengths of volumetric and ray-based metrics and persistent challenges in real-world instance-aware prediction. PanoBench is available at https://mias.group/PanoBench.

I. INTRODUCTION

Panoptic occupancy adds instance identities to semantic voxel labels, but existing real-world datasets offer incomplete geometry and synthetic-only data cannot capture real-world diversity. PanoBench unifies synthetic and reconstructed real-world annotations under a common panoptic evaluation protocol.

I. INTRODUCTION

  • Existing semantic occupancy labels lack instance identities, limiting evaluation of object separation, complete shapes, and temporal consistency.The gap is especially important for camera-centric scenes with partially observed objects that require multi-view and temporal cues.Source: §I · p. 1
  • Real-world datasets reconstructed from sparse, noisy, visibility-limited observations are difficult to extend to reliable instance-level supervision, while synthetic-only benchmarks omit real-world diversity.The paper motivates combining high-fidelity synthetic annotations with reconstructed real-world annotations under one task definition and evaluation protocol.Source: §I · p. 1
  • PanoBench combines a mesh-annotated synthetic dataset with two reconstructed real-world datasets for camera-centric panoptic occupancy prediction.The three subsets are PanoBench-CarlaOcc, PanoBench-nuScenes, and PanoBench-KITTI360.Source: §I · pp. 1–2
  • The benchmark supports unified study of semantic occupancy, panoptic occupancy, and 4D occupancy forecasting across synthetic and real-world driving scenes.Its canonical format and evaluation protocol are shared across the benchmark datasets.Source: Figure 1 · p. 2, §I · p. 2
  • For real-world data, PanoBench reconstructs static backgrounds and generates complete foreground instance meshes before fusing them into sparse panoptic voxels.The pipeline uses geometric priors and topology constraints for backgrounds and multi-view image observations for foreground assets.Source: §I · p. 2
  • Temporally aware valid masks retain complete thing instances supported by historical observations while evaluating stuff geometry within observable camera space.This protocol assesses object completeness and identity beyond surfaces visible in the target frame.Source: §I · p. 2

II. RELATED WORK

Prior occupancy work advances semantic scene completion and camera-based prediction, while panoptic methods add instance-aware 3D reasoning. PanoBench addresses the remaining scarcity of real-world panoptic datasets and defines a sparse voxel representation for prediction.

A. Semantic Occupancy Prediction Benchmarks

  • Real-world occupancy benchmarks improve scene coverage and label density, but their annotations remain incomplete because of sparse LiDAR and occlusion; simulation offers complete geometry but leaves a domain gap.PanoBench bridges simulated annotations with more realistic and diverse driving scenarios.Source: A. Semantic Occupancy Prediction… · p. 3

B. Vision-Based Occupancy Prediction

  • Vision-based occupancy methods lift image features into 3D representations or exploit sparse scene structures, while forecasting methods predict future occupancy from past observations.Existing supervision and evaluation remain dominated by semantic occupancy, leaving panoptic instance quality and cross-dataset generalization underexplored.Source: B. Vision-Based Occupancy Prediction · p. 3

C. Panoptic Occupancy Prediction

  • Panoptic occupancy extends semantic scene completion with instance-aware 3D prediction, but public panoptic datasets remain scarce and synthetic-only data cannot capture real-world characteristics.PanoBench systematizes CarlaOcc and adds reconstructed nuScenes and KITTI-360 subsets for real-world and cross-domain study.Source: C. Panoptic Occupancy Prediction · pp. 3–4

A. Task Formulation

  • The task estimates an ego-centric voxel-space panoptic scene from historical multi-view RGB observations and camera geometry.The canonical representation stores occupied voxels sparsely, with semantic categories and instance identities encoded per voxel.Source: A. Task Formulation · p. 4

B. PanoBench Overview

PanoBench shares a canonical occupancy form, valid-mask representation, and model interface across one synthetic and two real-world subsets. The synthetic subset derives complete scene geometry from simulated meshes, while real-world data use multi-modal observations and annotations.

B. PanoBench Overview

  • The benchmark family includes PanoBench-CarlaOcc, PanoBench-nuScenes, and PanoBench-KITTI360, with shared formats and interfaces for occupancy tasks and sim-to-real transfer.Subsets vary by domain, camera view, and voxel resolution.Source: B. PanoBench Overview · p. 4

C. PanoBench-CarlaOcc: A Synthetic Panoptic Occupancy Dataset

  • PanoBench-CarlaOcc contains 104 sequences and 104,000 frames from eight CARLA towns, with 30 semantic classes and persistent instance identities.It provides forward- and surround-view variants over a 51.2×51.2×6.4 m volume at 0.1, 0.2, and 0.4 m voxel resolutions.Source: B. PanoBench Overview · p. 4

B. PanoBench Overview

  • The real-world subsets provide 34,149 nuScenes frames and 76,473 KITTI-360 frames, with different camera settings, class counts, spatial extents, and voxel resolutions.They combine multi-view RGB images, semantic LiDAR, HD maps, and 3D object annotations, and include complete foreground assets with persistent instance labels.Source: B. PanoBench Overview · p. 4

1) Data Collection:

  • The synthetic data collection uses six synchronized cameras rendering RGB, depth, and semantic images alongside a 32-channel LiDAR at 20 Hz.Sequences cover varied road layouts, architectural styles, and four traffic-density levels.Source: 1) Data Collection · pp. 4–5

2) Mesh-Based Scene Reconstruction:

3) Panoptic Occupancy Generation:

PanoBench reconstructs real-world panoptic occupancy by combining background geometry and foreground instance meshes in a shared voxel space. Its background pipeline uses LiDAR and map priors, then repairs topology and filters non-planar structures.

D. Real-World Panoptic Occupancy Reconstruction

  • Sparse, noisy, incomplete observations constrain real-world reconstruction, while synthetic scenes have limited diversity and a persistent simulation-to-real domain gap.Conventional multi-frame aggregation can also introduce ghosting and fragmentation in sparsely observed distant regions.Source: D. Real-World Panoptic Occupancy… · p. 5
  • Real-world annotations integrate reconstructed background geometry and foreground instance meshes through three coordinated stages: background reconstruction, mesh acquisition, and occupancy generation.The stages combine complementary evidence to produce geometrically complete, temporally consistent, instance-aware annotations.Source: D. Real-World Panoptic Occupancy… · p. 5

1) Background Point Cloud Reconstruction:

  • Background reconstruction aggregates labeled static LiDAR points across a sequence and lifts HD-map polygons into ground layers using map regions as horizontal priors and LiDAR as vertical evidence.Foreground points are removed before aggregation to prevent motion-induced trailing artifacts.Source: 1) Background Point Cloud Reconstruction · p. 5
  • Topology-aware recovery identifies map-adjacent candidate flat regions, checks their normals, and uses shared boundaries with Thin-Plate Spline interpolation to restore continuous elevations.Propagation is limited to a buffer around known seeds to avoid filling distant unobserved terrain.Source: 1) Background Point Cloud Reconstruction · pp. 6–7

3) Panoptic Occupancy Generation:

  • The pipeline filters and locally refines non-planar background point clouds, then combines the recovered scene background with aligned foreground meshes before voxelization and fusion.Statistical outlier removal suppresses isolated noise and floating artifacts while preserving locally dense structures.Source: Figure 2 · p. 6, 1) Background Point Cloud Reconstruction · p. 7

2) Foreground Mesh Asset Acquisition:

PanoBench reconstructs complete foreground things from selected multi-frame, multi-view image crops rather than relying on sparse LiDAR surfaces. It builds a diverse prototype mesh pool and assigns a compatible mesh to each required instance.

2) Foreground Mesh Asset Acquisition:

  • Sparse, temporally aggregated LiDAR captures only visible surfaces and can distort dynamic objects, so foreground reconstruction is formulated as image-conditioned mesh generation.The pipeline selects clean, viewpoint-diverse crops using image quality, mask coverage, and multi-camera constraints.Source: 2) Foreground Mesh Asset Acquisition · p. 7
  • Selected crop-mask pairs feed a generative mesh model, with multi-view reconstruction used when possible and a single-view branch for rare single-view cases.Decoded meshes and Gaussians are baked into textured models and simplified while preserving silhouette fidelity.Source: 2) Foreground Mesh Asset Acquisition · p. 7
  • Farthest Mesh Sampling builds a compact prototype pool that removes near-duplicates while retaining semantic coverage and within-category shape diversity.Meshes are grouped by taxonomy, centered, normalized by bounding-box diagonal, and compared using symmetric Chamfer distance.Source: 2) Foreground Mesh Asset Acquisition · p. 7
  • Hierarchical assignment uses an instance’s own reconstruction when available, otherwise relaxing retrieval from the same subcategory to the semantic superclass.Among candidates, proportional shape compatibility is primary and absolute dimensional similarity breaks ties.Source: 2) Foreground Mesh Asset Acquisition · pp. 7–8

3) Panoptic Occupancy Ground Truth Generation:

Background points and assigned foreground meshes are aligned and voxelized into sparse panoptic labels, with elevation-aware priorities resolving overlaps. The valid-mask protocol evaluates historically supported whole thing instances, visible stuff voxels, and camera-visible free space.

3) Panoptic Occupancy Ground Truth Generation:

  • The background point cloud and assigned thing meshes are transformed into the ego frame and voxelized on a common lattice, with background identity zero and persistent identities for things.The two representations are assembled into per-frame sparse panoptic annotations.Source: 3) Panoptic Occupancy Ground Truth… · p. 8
  • Elevation-aware label fusion resolves competing background and foreground voxel labels using category priorities and world-space elevations, preserving higher-priority surfaces and foreground objects.The output stores occupied voxel coordinates, semantic categories, and persistent instance identities.Source: 3) Panoptic Occupancy Ground Truth… · p. 8

E. Valid Mask Generation

  • Frame-wise visibility masks miss objects observed earlier but currently occluded and can score only visible object fragments rather than complete instances.The proposed mask aggregates visibility over a temporal window and makes a whole instance valid once any of its voxels is observed.Source: E. Valid Mask Generation · pp. 8–9
  • Voxel visibility is tested by ray marching and aggregated across the temporal window; instance validity follows when any voxel in the instance is visible.This protocol evaluates complete things while retaining conservative, voxel-level treatment of stuff.Source: E. Valid Mask Generation · pp. 8–9
  • The temporal mask keeps historically supported thing instances, scores stuff at the voxel level, and limits free-space evaluation to camera-observable rays.Stuff remains valid only where visible, and foreground things do not invalidate background stuff behind or beneath them.Source: E. Valid Mask Generation · p. 9

IV. EXPERIMENTS

The experiments compare PanoBench annotations with existing benchmarks across nuScenes, KITTI-360, and CARLA. Camera-centric methods receive RGB images and camera and ego-pose parameters, and are evaluated on three occupancy tasks.

1) Datasets and Benchmarks:

  • The benchmark evaluation covers PanoBench-nuScenes, PanoBench-KITTI360, and PanoBench-CarlaOcc against public occupancy benchmarks in nuScenes, KITTI-360, and CARLA.Methods receive RGB images, camera intrinsics and extrinsics, and ego poses during training and inference.Source: 1) Datasets and Benchmarks · p. 9
  • On nuScenes, the six-camera surround-view protocol uses 0.4 m voxels over an 80 × 80 × 6.4 m volume and retains official scene-level training and validation splits.PanoBench-nuScenes is compared with SurroundOcc, Occ3D, and OpenOccupancy.Source: 1) Datasets and Benchmarks · p. 9
  • On KITTI-360, the forward-view protocol uses the front camera, 0.2 m voxels, and official sequence-level partitions to prevent temporally adjacent frames crossing splits.The comparison is with SSCBench-KITTI360.Source: 1) Datasets and Benchmarks · p. 9
  • On CARLA, the six-camera protocol uses 0.4 m voxels and the official CarlaOcc split, with comparisons to CarlaSC and reconstructed Occ3D-style and SurroundOcc-style labels for the mesh-oracle study.The 0.1 m and 0.2 m variants are reserved for multi-resolution analysis.Source: 1) Datasets and Benchmarks · p. 9
  • Each PanoBench subset is evaluated on semantic occupancy prediction, panoptic occupancy prediction, and 4D occupancy forecasting at 1-, 2-, and 3-second horizons.Source: 1) Datasets and Benchmarks · p. 9

2) Benchmark Details and Metrics:

The section specifies a shared benchmark and metric setup, then evaluates annotation quality through observation fidelity and geometric completeness. The comparisons report higher camera-view agreement and more complete geometry for PanoBench annotations.

2) Benchmark Details and Metrics:

  • The semantic benchmark covers GaussianFormer, OPUS, SparseOcc, and Panoptic-FlashOcc; panoptic evaluation includes instance predictions from SparseOcc and Panoptic-FlashOcc, while forecasting uses SparseWorld and SparseWorld-TC.Methods share subset settings and optimization budget except for input-frame count; metrics include geometric IoU, semantic mIoU, and ray-based metrics under the valid-mask protocol.Source: 2) Benchmark Details and Metrics · p. 9

1) Occupancy Observation Fidelity:

  • PanoBench-nuScenes gains 7.2 points over OpenOccupancy at 0.2 m and 3.8 over Occ3D at 0.4 m; PanoBench-KITTI360 reaches 81.5% PixAcc at 0.2 m, 11.6 points above SSCBench-KITTI360.PixAcc compares rendered occupancy semantics with 2D semantic ground truth, restricted to pixels whose 3D points lie within the occupancy ground-truth range.Source: Table II · p. 10, 1) Occupancy Observation Fidelity · p. 10

B. Occupancy Annotation Quality

  • Annotation quality is assessed along observation fidelity, geometric completeness, spatial continuity, and temporal consistency; completeness concerns recovering full structures and silhouettes, including occluded or sparsely observed regions.Cross-dataset comparisons align region of interest, semantic taxonomy, and voxel resolution.Source: B. Occupancy Annotation Quality · p. 10, 2) Geometric Completeness · p. 10

2) Geometric Completeness:

  • On nuScenes, SurroundOcc, OpenOccupancy, and Occ3D show broken ground layers and incomplete foreground objects, while PanoBench yields more continuous backgrounds and complete shapes.The controlled Town10HD visualization shows PanoBench recovering continuous ground and complete vehicles close to the mesh oracle; denser Poisson-reconstructed surfaces still leave errors in complex structures.Source: 2) Geometric Completeness · p. 10
  • 58.9 IoU and 52.0 mIoU: PanoBench exceeds the stronger Occ3D-style pipeline by 20.6 and 19.4 points, respectively, against the mesh oracle.The largest reported gains are on Vehicle, Pedestrian, Terrain, and Sidewalk; Building and Vegetation remain difficult across pipelines.Source: 2) Geometric Completeness · p. 10

3) Spatial Continuity:

Spatial continuity is measured by the prevalence of small fragmented components, and PanoBench reports lower floating-voxel ratios across synthetic and real-world data.

3) Spatial Continuity:

  • Spatial continuity groups same-class occupied voxels into 26-connected components and treats a component as floating when its volume is at or below threshold V.The floating-voxel ratio rfp is reported at thresholds of 0.1 and 0.2 m3.Source: 3) Spatial Continuity · pp. 10–11
  • The reported floating-voxel ratios are compared at matched voxel sizes, with lower ratios indicating fewer small fragmented components.The table reports ratios in permille and marks the better result at each voxel size.Source: 3) Spatial Continuity · p. 11, Table IV · p. 12
  • PanoBench lowers floating-voxel ratios by 84.1% and 76.7% versus CarlaSC on CARLA, and by 63.5% on nuScenes and 46.5% on KITTI-360 versus counterparts.The CARLA reductions correspond to rfp@0.1 and rfp@0.2 averaged over two voxel sizes; real-world reductions average both thresholds at matched voxel sizes.Source: Table IV · p. 12, 3) Spatial Continuity · p. 12

4) Temporal Consistency:

PanoBench annotations are more temporally consistent than existing benchmarks in nearly all matched-voxel comparisons, with stable background geometry and foreground instance identities. Ablations show gains from map-guided background completion and view-selected foreground meshes, while semantic occupancy results expose a completeness–precision trade-off.

4) Temporal Consistency:

  • 12.5 and 12.1 mIoU-point gains and 5.6 and 5.5 IoU-point gains over CarlaSC at 0.2 and 0.4 m accompany temporally coherent background and foreground annotations.On nuScenes, PanoBench-nuScenes has the highest mIoU at both voxel sizes; on KITTI-360, it exceeds SSCBench-KITTI360 by 6.8 mIoU and 5.9 IoU points at 0.2 m.Source: 4) Temporal Consistency · p. 12
  • On nuScenes, PanoBench-nuScenes improves IoU over Occ3D by 7.2 points at 0.4 m, while its 0.2 m IoU is comparable to OpenOccupancy; coarser voxels yield higher consistency across datasets.SurroundOcc has the highest nuScenes IoU, 74.3.Source: 4) Temporal Consistency · p. 12

C. Ablation Study

  • 10.9 points of PixAcc improvement over semantic LiDAR aggregation come from the full background pipeline, with ground lifting and topology-aware recovery providing the main gains.Ground lifting raises PixAcc from 68.5% to 74.1%; flat-region recovery adds 4.9 points, while non-planar refinement adds 0.4.Source: C. Ablation Study · p. 12, Table VI · p. 13
  • 67.2% instance IoU is achieved by view-selected multi-view reconstruction, 2.1 points above using every eligible view and 6.0 points above the best single-view model.LiDAR aggregation reaches 12.4% instance IoU, and SAM 3D is the single-view fallback when only one mined observation survives.Source: C. Ablation Study · p. 13, Table VII · p. 13

1) Semantic Occupancy Prediction:

  • GaussianFormer leads IoU across all subsets, but SparseOcc surpasses it at every KITTI-360 RayIoU tolerance despite having about half its IoU.The results distinguish volumetric completeness from first-hit surface precision: GaussianFormer fills occluded volume but can extend occupancy in front of surfaces, while SparseOcc leaves more occluded volume empty.Source: 1) Semantic Occupancy Prediction · pp. 13–14, Table VIII · p. 14
  • Semantic occupancy methods also differ by domain: Panoptic-FlashOcc ranks second in CarlaOcc RayIoU@1 but last or tied for last on both real-world subsets.PanoBench-nuScenes has the highest mIoU at both voxel sizes; PanoBench-KITTI360 exceeds SSCBench-KITTI360 by 6.8 mIoU and 5.9 IoU points at 0.2 m.Source: 4) Temporal Consistency · p. 12, 1) Semantic Occupancy Prediction · p. 13

2) Panoptic Occupancy Prediction:

Panoptic results show that instance quality does not reliably track semantic or volumetric accuracy, while forecasting performance declines with longer horizons. Across tasks, metric rankings vary and real-world instance grouping remains difficult.

2) Panoptic Occupancy Prediction:

  • SparseOcc has higher RayPQ at every depth tolerance on all three subsets; on nuScenes, its RayPQ@1 is over three times Panoptic-FlashOcc’s despite FlashOcc’s higher IoU.For both methods, RayPQ@4 retains about two thirds or more of RayIoU@4 on CarlaOcc, but about half or less on the real-world subsets.Source: 2) Panoptic Occupancy Prediction · p. 14
  • On nuScenes, relaxing RayPQ tolerance from 1 m to 4 m more than triples Panoptic-FlashOcc’s score but less than doubles SparseOcc’s, suggesting depth displacement of FlashOcc instances.Panoptic-FlashOcc groups thing voxels around BEV instance centers after depth-based lifting, whereas SparseOcc predicts instance masks directly in 3D.Source: 2) Panoptic Occupancy Prediction · p. 14

3) 4D Occupancy Forecasting:

  • Both forecasting methods degrade monotonically with horizon, with semantics falling faster than geometry; KITTI-360 loses more than a third of its 1 s mIoU by 3 s.The decline is mildest on nuScenes, whose surround-view volume remains largely within observed regions; IoU is dominated by static background, while mIoU weights classes equally and is sensitive to small dynamic objects.Source: 3) 4D Occupancy Forecasting · pp. 14–15, Table X · p. 15
  • SparseWorld-TC outperforms SparseWorld at every horizon on CarlaOcc and KITTI-360, but the methods are nearly indistinguishable on nuScenes.On KITTI-360, SparseWorld-TC’s average mIoU and IoU are higher by roughly a quarter and a fifth, respectively.Source: 3) 4D Occupancy Forecasting · p. 15, Table X · p. 15

4) Discussion:

  • No single architecture dominates across metrics, synthetic rankings do not reliably transfer to real data, and limited camera coverage hampers current and future prediction.The discussion identifies instance grouping, rather than semantic labeling, as the main bottleneck in real scenes and notes that depth errors favor instance reasoning directly in 3D.Source: 4) Discussion · p. 15

1) Multi-Resolution Analysis:

Resolution, domain transfer, and storage analyses show both practical benefits and constraints in using PanoBench. The paper concludes that occupancy methods remain uneven across metrics and real-world instance-aware prediction remains challenging.

1) Multi-Resolution Analysis:

  • Reducing CarlaOcc voxel size from 0.4 m to 0.2 m lowers average IoU and mIoU by 11.5 and 9.3 points; reducing it further to 0.1 m lowers them by 23.7 and 20.1 points.Finer grids multiply cell counts by eight at each halving, while image resolution and depth cues remain unchanged, making localization errors and thin structures more consequential.Source: 1) Multi-Resolution Analysis · p. 15, Table XI · p. 15

2) Synthetic-to-Real Transfer Learning:

  • CarlaOcc pretraining consistently improves mIoU and RayIoU for both architectures on both target domains, with larger gains on nuScenes than KITTI-360.On nuScenes, SparseOcc gains 1.2 mIoU, 1.7 RayIoU, and 1.5 RayPQ points; on KITTI-360, SparseOcc’s RayPQ decreases by 0.2 points despite its mIoU and RayIoU gains.Source: 2) Synthetic-to-Real Transfer Learning · pp. 15–16, Table XII · p. 15

3) Canonical Storage and I/O Efficiency:

  • The canonical sparse format cuts storage by more than 95% and per-frame latency by more than 83% for dense releases, and roughly halves Occ3D storage while reducing its loading latency by more than 85%.The authors attribute the latency reduction largely to smaller serialized payloads requiring fewer bytes to be read and decoded.Source: 3) Canonical Storage and I/O Efficiency · p. 16, Table XIII · p. 16

V. CONCLUSION AND FUTURE WORK

  • PanoBench combines complete mesh-derived synthetic supervision with reconstructed real-world background geometry and foreground assets under a common sparse representation, valid masks, and persistent instance identities.The authors identify longer-horizon 4D panoptic forecasting, open-set categories, and multimodal supervision with camera-only inference as future work.Source: §V · p. 16
Loading 2610.23537v1…