WayPOP: A Panoramic Open-Set Panoptic Tracking Benchmark

Dataset Summary

WayPOP is a benchmark for panoramic open-set panoptic segmentation and tracking in autonomous driving. It evaluates whether perception systems can segment known classes and unknown objects while maintaining consistent instance identities across time and camera views.

The dataset augments real driving sequences from the Waymo Panoramic Video Panoptic Segmentation (PVPS) dataset with curated 3D objects from Objaverse 1.0 and Objaverse-XL. Inserted objects fall outside the benchmark’s known semantic taxonomy and are labeled as unknown. They are placed at fixed locations in the 3D scene and rendered consistently across synchronized camera views and time, with corresponding updates to LiDAR measurements.

The release provides RGB images from five synchronized cameras, vehicle-frame XYZ point clouds, camera calibration, timestamps, and poses. Validation images include panoptic annotations with consistent instance identities across time and views. Test annotations are withheld for evaluation through the benchmark server.

See the accompanying WayPOP paper for dataset construction, task definitions, and the evaluation protocol.

Benchmark Evaluation

Evaluation will be hosted on Codabench through three task-specific competitions:

• Multi-View Open-Set Panoptic Tracking (MV-OSPT): Panoptic segmentation with consistent instance identities across time and all five camera views.
• Frontal-View Open-Set Panoptic Tracking (FV-OSPT): Open-set panoptic tracking in the front-facing camera, maintaining consistent instance identities over time.
• View-Independent Open-Set Panoptic Segmentation (VI-OSPS): Open-set panoptic segmentation of individual images without temporal or cross-camera association.

All three tasks evaluate both known classes and unknown objects.

Access and Terms

To use the Waymo-derived data, register at Waymo Open Dataset and accept the Waymo Dataset License Agreement for Non-Commercial Use. The inserted 3D assets retain their individual source licenses. Read TERMS_OF_USE before using or sharing WayPOP; LICENSE provides attribution and upstream-license details. WayPOP is not affiliated with Waymo LLC.

Splits

The paper calls the two test splits WayPOP-v1 and WayPOP-XL. Their short IDs are used in folder names, manifests, archives, and download commands.

Public name	Split ID	Description
Validation	val	Six Objaverse 1.0 sequence variants with public panoptic annotations.
WayPOP-v1	v1	Test split with Objaverse 1.0 objects and private annotations.
WayPOP-XL	xl	Test split with Objaverse-XL objects and private annotations.

Dataset Structure

The release contains one metadata archive and one archive per sequence variant. After extraction, files follow the layout below. Data directories use the split IDs defined in Splits.

waypop/
├── README.md
├── FORMAT.md
├── LICENSE
├── TERMS_OF_USE.md
├── categories.json
├── release.json
├── sequences.jsonl
├── SHA256SUMS
├── assets/
│   └── teaser.gif
├── tools/
│   ├── read_sample.py
│   └── verify_download.py
├── val/
│   ├── images/
│   ├── pointcloud/
│   ├── panoptic/
│   ├── wstq_meta/
│   └── meta/
│       ├── samples.jsonl
│       └── sequences/
├── v1/
│   ├── images/
│   ├── pointcloud/
│   └── meta/
└── xl/
    ├── images/
    ├── pointcloud/
    └── meta/

File Organization

RGB images (images/) and panoptic annotations (panoptic/) are organized by sequence, camera, and frame. Point clouds (pointcloud/) are organized by sequence and frame. All RGB images are stored as PNG files.

Sample Metadata

Each data directory contains a meta/samples.jsonl file with one record per camera-frame sample. Each record includes:

• Identifiers: Sample and sequence identifiers, source sequence, clip, and asset source.
• Camera and frame information: Camera information, frame index, timestamp, and image dimensions.
• Data paths: Paths to the corresponding RGB image and point cloud.
• Poses: Vehicle and image poses.

Validation records additionally include paths to panoptic annotations and wSTQ metadata, together with the panoptic label divisor.

Download

After access to the WayPOP dataset repository is granted, download and extract the complete release:

python -m pip install -U huggingface_hub
hf auth login
hf download mohangrim/WayPOP --repo-type dataset --local-dir waypop_download
python waypop_download/reconstruct_release.py waypop_download waypop
python waypop/tools/verify_download.py waypop

To download only validation data:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="mohangrim/WayPOP",
    repo_type="dataset",
    local_dir="waypop_download",
    allow_patterns=[
        "archive_index.json",
        "reconstruct_release.py",
        "README.md",
        "FORMAT.md",
        "TERMS_OF_USE.md",
        "waypop-2026-metadata.tar",
        "waypop-2026-val-*.tar",
    ],
)

python waypop_download/reconstruct_release.py waypop_download waypop --split val
python waypop/tools/verify_download.py waypop --split val

The extraction script checks archive hashes before writing files. Both tools also accept --sequence SEQUENCE_ID for an individual variant. Markdown documentation is copied from the download folder and can be updated without changing the checksummed data archives.

Reading the Data

See FORMAT.md for sample paths, sensor metadata, validation annotations, and benchmark prediction and submission formats.

Classes

categories.json is the authoritative taxonomy and benchmark category list. Predictions use these raw category IDs:

Raw ID	Category	Type
2	car	thing
3	truck	thing
4	bus	thing
5	other_large_vehicle	thing
6	bicycle	stuff
7	motorcycle	stuff
8	trailer	thing
9	pedestrian	thing
10	cyclist	thing
11	motorcyclist	stuff
14	construction_cone_pole	stuff
15	pole	stuff
16	pedestrian_object	stuff
17	sign	stuff
18	traffic_light	stuff
19	building	stuff
20	road	stuff
21	lane_marker	stuff
22	road_marker	stuff
23	sidewalk	stuff
24	vegetation	stuff
25	sky	stuff
29	unknown	unknown/OOD

Citation

If you use WayPOP in your research, please cite our paper:

WayPOP: A Panoramic Open-Set Panoptic Tracking Benchmark (NeurIPS, 2026)

@inproceedings{waypop2026,
  title={WayPOP: A Panoramic Open-Set Panoptic Tracking Benchmark},
  author={Mohan, Rohit and Luz, Maximilian and Chowdhury, Swarnava and Lu, Yao and Drews, Florian and N{\"u}rnberg, Thomas and Miron, Yakov and Tombari, Federico and Gasperini, Stefano and Valada, Abhinav},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}

Contact

For dataset questions, bug reports, benchmark submission issues, or collaboration inquiries, contact Rohit Mohan at mohan@cs.uni-freiburg.de.