PanoBench: A Panoptic Occupancy Prediction Benchmark for Autonomous Driving
Paper • 2610.23537 • Published
PanoBench model checkpoints
Code · Paper · Dataset · Project
Fourteen checkpoints for semantic occupancy prediction, panoptic occupancy prediction and 4D occupancy forecasting. Checkpoint configurations, subsets, voxel variants, file sizes and download links are listed in the code README.
OPUS-V1 follows the official occupied-class classifier design. SparseOcc and FlashOCC support panoptic evaluation; SparseWorld-TC predicts future occupancy.
Download and evaluate
Install the optional download CLI with pip install huggingface_hub. From the cloned code repository:
hf download fengyi233/PanoBench \
opusv1_nuscenes_vs_0.4.pth \
--local-dir checkpoints
bash scripts/eval.sh experiments/opusv1/panobench_nuscenes \
checkpoints/opusv1_nuscenes_vs_0.4.pth vs_0.4
Follow the installation, data and evaluation guides. These checkpoints are for the PanoBench codebase and its model-specific OpenMMLab/CUDA components.
Measured validation results
All values below are percentages; higher is better. Results come from completed full-validation runs for the linked checkpoint. Paper-reported numbers are in the paper.
Forecasting averages cover the 1 s, 2 s and 3 s horizons; the current frame is reported separately. All evaluations use the dataset's valid mask.
KITTI360 forecasting scope: the release provides 64,418 labels and 12,782 validation samples with complete future labels. The historical KITTI360 forecasting measurements below used a larger future-label set. No rerun on the released forecasting subset is reported.
FlashOCC · PanoBench-CarlaOcc · vs_0.2_surround_view
- IoU: 62.80
- mIoU: 32.06
- SQ: 54.96
- RQ: 17.86
- PQ: 12.58
- RayPQ@1m: 21.22
- RayPQ@2m: 24.35
- RayPQ@4m: 25.70
FlashOCC · PanoBench-KITTI360 · 0.2 m
Validation samples: 13,164.
- IoU: 41.17
- mIoU: 11.35
- RayIoU@1m: 18.79
- RayIoU@2m: 25.02
- RayIoU@4m: 27.58
- SQ: 38.55
- RQ: 4.41
- PQ: 2.64
- RayPQ@1m: 7.05
- RayPQ@2m: 10.73
- RayPQ@4m: 12.58
FlashOCC · PanoBench-nuScenes · vs_0.4
Validation samples: 6,019.
- IoU: 47.91
- mIoU: 22.80
- RayIoU@1m: 25.17
- RayIoU@2m: 34.87
- RayIoU@4m: 41.12
- SQ: 52.40
- RQ: 7.12
- PQ: 4.31
- RayPQ@1m: 3.25
- RayPQ@2m: 7.05
- RayPQ@4m: 10.84
GaussianFormer-V1 · PanoBench-CarlaOcc · vs_0.4_surround_view
Validation samples: 12,000.
- IoU: 82.23
- mIoU: 46.21
- RayIoU@1m: 56.42
- RayIoU@2m: 60.97
- RayIoU@4m: 62.78
GaussianFormer-V1 · PanoBench-KITTI360 · 0.2 m
Validation samples: 13,164.
- IoU: 65.26
- mIoU: 15.62
- RayIoU@1m: 21.10
- RayIoU@2m: 25.51
- RayIoU@4m: 27.14
GaussianFormer-V1 · PanoBench-nuScenes · vs_0.4
Validation samples: 6,019.
- IoU: 66.50
- mIoU: 32.64
- RayIoU@1m: 32.80
- RayIoU@2m: 41.77
- RayIoU@4m: 46.64
OPUS-V1 · PanoBench-CarlaOcc · vs_0.4_surround_view
Validation samples: 12,000.
- IoU: 73.45
- mIoU: 41.22
- RayIoU@1m: 50.88
- RayIoU@2m: 57.41
- RayIoU@4m: 59.85
OPUS-V1 · PanoBench-nuScenes · vs_0.4
Validation samples: 6,019.
- IoU: 50.06
- mIoU: 29.91
- RayIoU@1m: 32.63
- RayIoU@2m: 41.39
- RayIoU@4m: 46.22
SparseOcc · PanoBench-CarlaOcc · vs_0.4_surround_view
Validation samples: 12,000.
- IoU: 68.53
- mIoU: 41.72
- RayIoU@1m: 50.06
- RayIoU@2m: 56.67
- RayIoU@4m: 60.11
- SQ: 68.38
- RQ: 38.18
- PQ: 27.50
- RayPQ@1m: 36.51
- RayPQ@2m: 41.80
- RayPQ@4m: 44.17
SparseOcc · PanoBench-KITTI360 · 0.2 m
Validation samples: 13,164.
- IoU: 33.28
- mIoU: 11.17
- RayIoU@1m: 28.06
- RayIoU@2m: 32.01
- RayIoU@4m: 33.36
- SQ: 47.47
- RQ: 4.13
- PQ: 2.46
- RayPQ@1m: 12.65
- RayPQ@2m: 15.97
- RayPQ@4m: 17.40
SparseOcc · PanoBench-nuScenes · vs_0.4
Validation samples: 6,019.
- IoU: 55.01
- mIoU: 32.96
- RayIoU@1m: 33.74
- RayIoU@2m: 41.78
- RayIoU@4m: 46.53
- SQ: 57.21
- RQ: 17.90
- PQ: 11.30
- RayPQ@1m: 11.11
- RayPQ@2m: 16.42
- RayPQ@4m: 20.12
SparseWorld-TC · PanoBench-CarlaOcc · vs_0.4_surround_view
Validation samples: 10,140.
- IoU_0s: 73.20
- mIoU_0s: 40.35
- IoU_1s: 69.25
- mIoU_1s: 34.45
- IoU_2s: 64.81
- mIoU_2s: 29.84
- IoU_3s: 60.40
- mIoU_3s: 25.72
- IoU_avg: 64.82
- mIoU_avg: 30.01
SparseWorld-TC · PanoBench-KITTI360 · 0.2 m
Measured on the historical forecasting split, which used future labels beyond the released 64,418-frame list. These metrics are not a validation of the released 12,782-sample forecasting split.
- IoU_0s: 32.25
- mIoU_0s: 9.63
- IoU_1s: 28.63
- mIoU_1s: 7.66
- IoU_2s: 24.78
- mIoU_2s: 6.06
- IoU_3s: 21.10
- mIoU_3s: 4.79
- IoU_avg: 24.84
- mIoU_avg: 6.17
SparseWorld-TC · PanoBench-nuScenes · vs_0.4
- IoU_0s: 50.67
- mIoU_0s: 27.48
- IoU_1s: 49.24
- mIoU_1s: 24.87
- IoU_2s: 47.64
- mIoU_2s: 22.72
- IoU_3s: 45.92
- mIoU_3s: 20.84
- IoU_avg: 47.60
- mIoU_avg: 22.81
Citation
Please cite PanoBench when using these checkpoints. The citation is available in the code repository.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support