PanoBench

Code · Paper · Models · Project

PanoBench supports semantic occupancy prediction, panoptic occupancy prediction, and 4D occupancy forecasting across three complementary driving subsets. Downloads under panobench/ mirror the local data/ directory. They include occupancy labels, preprocessed CarlaOcc and KITTI360 images, calibration, poses, frame splits and nuScenes sweep information. KITTI360 benchmark frame lists, class definitions and mappings are included in the code repository. Obtain nuScenes images from the official dataset.

Repository layout

• panobench/canonical_occ/: nine label archives, with CarlaOcc and nuScenes grouped by subset and variant.
• panobench/CarlaOccV1Preprocessed/: 104 sequence archives containing six-camera preprocessed RGB images and poses, plus calib.tar and splits.tar.
• panobench/KITTI-360/: nine image_00 sequence archives under data_2d_raw/, plus data_poses.tar and calibration.tar.
• panobench/nuscenes/pkl_files_panobench/: training and validation sweep information PKLs.
• public_occupancy_datasets/: existing canonical-format conversions of Occ3D-nuScenes, OpenOccupancy, SurroundOcc, SSCBench-KITTI360 and SemanticKITTI.

Subsets and splits

• PanoBench-CarlaOcc: 104,000 frames per variant: 80,000 train, 12,000 validation and 12,000 test. All three splits are provided. Six variants combine 0.1 / 0.2 / 0.4 m voxels with forward-view or surround-view. Forward-view uses only image_00; surround-view uses six cameras.
• PanoBench-nuScenes: 34,149 frames per variant: 28,130 train and 6,019 validation. Two variants use 0.2 m and 0.4 m voxels.
• PanoBench-KITTI360: 64,418 frames at 0.2 m: 51,254 train and 13,164 validation. Drive 0009 is the validation split. Exact frame lists are included in the code repository.

Together these represent 202,567 distinct subset/frame pairs and 756,716 label files across all nine voxel/view variants. Repeated frames at different resolutions are counted separately only in the label-file total.

KITTI360 forecasting requires every target at raw frame offsets [0, 10, 20, 30] to occur in the released split list. This yields 49,299 training and 12,782 validation samples with complete horizons. Historical forecasting metrics used a larger future-label set and do not measure this restricted validation split.

Download and extract

Install the downloader with pip install huggingface_hub. Run all extraction commands from the PanoBench code repository root: every archive contains its complete data/... destination path. The downloader preserves remote paths under --local-dir downloads; downloaded files therefore appear in downloads/panobench/.

Occupancy labels

Labels are stored in panobench/canonical_occ/. Each voxel/view variant is one complete .tar archive. KITTI360 has no variant directory and uses a single panobench_kitti360.tar.

For example, download CarlaOcc 0.4 m surround-view labels:

hf download fengyi233/PanoBench \
  panobench/canonical_occ/panobench_carlaocc/vs_0.4_surround_view.tar \
  --repo-type dataset --local-dir downloads

tar -xf downloads/panobench/canonical_occ/panobench_carlaocc/vs_0.4_surround_view.tar

This extracts labels directly into data/canonical_occ/panobench_carlaocc/vs_0.4_surround_view/. KITTI360 benchmark split files are already included in datasets/splits/panobench_kitti360/ in the cloned code repository.

Use these archive paths for other subsets:

• CarlaOcc: panobench/canonical_occ/panobench_carlaocc/vs_0.1_forward_view.tar, or replace the resolution/view with another of the six supported combinations.
• nuScenes: panobench/canonical_occ/panobench_nuscenes/vs_0.2.tar or panobench/canonical_occ/panobench_nuscenes/vs_0.4.tar.
• KITTI360: panobench/canonical_occ/panobench_kitti360.tar.

Use the same archive filename in the download and extraction commands. See the Hugging Face download guide for alternative download methods.

CarlaOcc images, calibration, splits and poses

The panobench/CarlaOccV1Preprocessed/ directory contains 104 sequence archives, named <scene>.tar.gz. Each contains six cameras of 640 × 192 RGB images and that sequence's poses/cam_00.txt and poses/lidar.txt. The separate calib.tar and splits.tar provide the official CarlaOcc calibration and train/validation/test frame lists. All 106 files together occupy approximately 122.84 GB.

hf download fengyi233/PanoBench \
  --repo-type dataset --local-dir downloads \
  --include "panobench/CarlaOccV1Preprocessed/*.tar*"

for archive in downloads/panobench/CarlaOccV1Preprocessed/*.tar.gz; do
  tar -xzf "$archive"
done
tar -xf downloads/panobench/CarlaOccV1Preprocessed/calib.tar
tar -xf downloads/panobench/CarlaOccV1Preprocessed/splits.tar

To download an individual sequence, specify its path, for example panobench/CarlaOccV1Preprocessed/Town01_Opt_Seq00.tar.gz. The image archives are shared by all six label variants.

nuScenes sweep information

Download the nuScenes sweep information files for training and validation, and place them in data/nuscenes/pkl_files_panobench/:

hf download fengyi233/PanoBench \
  panobench/nuscenes/pkl_files_panobench/nuscenes_infos_train_sweep.pkl \
  panobench/nuscenes/pkl_files_panobench/nuscenes_infos_val_sweep.pkl \
  --repo-type dataset --local-dir downloads

mkdir -p data/nuscenes/pkl_files_panobench
cp downloads/panobench/nuscenes/pkl_files_panobench/nuscenes_infos_{train,val}_sweep.pkl data/nuscenes/pkl_files_panobench/

KITTI360 images, calibration and poses

The panobench/KITTI-360/ directory contains nine image archives under data_2d_raw/<scene>.tar.gz, plus data_poses.tar and calibration.tar. The image archives contain only 640 × 192 image_00 PNGs, including temporal context frames. data_poses.tar provides the nine data/KITTI-360/data_poses/<scene>/cam0_to_world_complete.txt files; calibration.tar provides the official calibration files. Images and supporting archives together occupy approximately 16.37 GB.

hf download fengyi233/PanoBench \
  --repo-type dataset --local-dir downloads \
  --include "panobench/KITTI-360/data_2d_raw/*.tar.gz" \
  --include "panobench/KITTI-360/*.tar"

for archive in downloads/panobench/KITTI-360/data_2d_raw/*.tar.gz; do
  tar -xzf "$archive"
done
tar -xf downloads/panobench/KITTI-360/data_poses.tar
tar -xf downloads/panobench/KITTI-360/calibration.tar

To download an individual sequence, specify its path, for example panobench/KITTI-360/data_2d_raw/2013_05_28_drive_0000_sync.tar.gz. Use the benchmark frame lists, class definitions and label mappings included in this code repository.

Label format

Each .npz has exactly five fields:

• occupancy: sparse (x, y, z, 1000 * semantic_id + instance_id) integer rows.
• valid_mask: packed dense validity bits in NumPy's default bit order.
• volume_size: (X, Y, Z) grid dimensions.
• voxel_size: voxel edge length in meters.
• voxel_origin: grid origin in meters.

import numpy as np

with np.load("FRAME.npz", allow_pickle=False) as frame:
    occupancy = frame["occupancy"]
    shape = tuple(int(v) for v in frame["volume_size"])
    valid = np.unpackbits(frame["valid_mask"], count=int(np.prod(shape)))
    valid = valid.reshape(shape).astype(bool)
    semantic_ids = occupancy[:, 3] // 1000
    instance_ids = occupancy[:, 3] % 1000

Semantic classes are subset-specific. Use the class metadata in utils/labels.py and evaluate within the released validity mask.

Images

The CarlaOcc and KITTI360 archives provide 640 × 192 preprocessed images ready for training and evaluation. Obtain nuScenes images from the official dataset. See the data guide for the local directory layout and optional preparation from original images.

Citation

Please cite PanoBench and CarlaOcc as appropriate; BibTeX is provided in the code repository.