mohangrim/panoptic-waymo

Panoptic Waymo

Dataset Summary

Panoptic Waymo is a high-resolution LiDAR panoptic segmentation benchmark derived from the Waymo Open Dataset. It provides panoptic annotations for the original Waymo Open Dataset training and validation splits, covering 798 training scenes and 202 validation scenes with 15 stuff classes and 6 thing classes.

The benchmark is designed for fine-grained 3D scene understanding and multi-modal perception, leveraging Waymo’s dense 64-beam LiDAR and five high-resolution cameras covering more than 180° of the scene. Given a LiDAR point cloud, the task is to predict a panoptic label for every point: a semantic class for all valid points and a temporally local instance ID for points belonging to thing classes. Stuff classes are represented only by their semantic class.

Dataset Access and Terms of Use

Panoptic Waymo contains annotations derived for LiDAR panoptic segmentation from the Waymo Open Dataset. Users must download the original Waymo Open Dataset separately from the official Waymo website and comply with the Waymo Dataset License Agreement for Non-Commercial Use.

Use of this dataset, the provided annotations, models trained or evaluated on this benchmark, and derivative works based on it is subject to the applicable Waymo Open Dataset terms.

Before downloading or using Panoptic Waymo, users must review and comply with:

https://waymo.com/open/terms

This dataset is not affiliated with, sponsored by, or endorsed by Waymo LLC.

Dataset Structure

The released label root has the following structure:

panoptic_waymo_labels/
  panoptic_waymo_label_mapping.yaml
  manifests/
    train.jsonl
    val.jsonl
  checksums/
    train.sha256
    val.sha256
  labels/
    training/
      <segment_id>/
        panoptic/
          <frame_id>.npz
    validation/
      <segment_id>/
        panoptic/
          <frame_id>.npz

Each manifest row describes one frame. Important fields are:

split          training or validation
segment_id     Waymo segment directory/name
frame_id       sparse original Waymo frame index, for example 000025
lidar_token    <frame_id>_<segment_id>
label_relpath  path to the released panoptic label file
point_count    number of LiDAR points in the frame

Frame IDs follow the original sparse Waymo TFRecord frame counters.

Getting the Full Dataset: Requirements and Preparation

The Panoptic Waymo release does not redistribute raw LiDAR scans, camera images, calibration files, or the original Waymo TFRecords. Users must therefore construct the full dataset locally using the official Waymo Open Dataset files and the Panoptic Waymo label release.

Data Requirements

To get started, you need two components:

1. The Panoptic Waymo label release from this repository.
2. The official Waymo Open Dataset v1.4.3 TFRecords, downloaded separately from Waymo.

Download the Panoptic Waymo labels with the Hugging Face CLI:

python -m pip install -U "huggingface_hub"
hf download mohangrim/panoptic-waymo \
  --repo-type dataset \
  --local-dir /path/to/panoptic_waymo_labels

Alternatively, download the dataset with Git LFS:

git lfs install
git clone https://huggingface.co/datasets/mohangrim/panoptic-waymo /path/to/panoptic_waymo_labels

The Waymo root directory is expected to contain training/ and validation/ subdirectories with files named in the following format:

<segment_id>_with_camera_labels.tfrecord

Panoptic Waymo Devkit

To construct the full dataset for LiDAR panoptic segmentation, use the Panoptic Waymo Devkit:

https://github.com/mohangrim/panoptic-waymo-devkit

The devkit provides the official tools for preparing and working with Panoptic Waymo. After downloading both required inputs, the devkit can reconstruct the local sensor files from the Waymo TFRecords, align them with the Panoptic Waymo labels, and provide utilities for visualization and evaluation.

Annotation Format

Each ground-truth label file is a compressed NumPy .npz file with key data:

labels = np.load(label_path)["data"]

The array is one-dimensional, has dtype uint32, and contains one panoptic label per LiDAR point in the corresponding Waymo point cloud.

Panoptic labels are encoded as:

panoptic_id = semantic_id * 1000 + instance_id
semantic_id = panoptic_id // 1000
instance_id = panoptic_id % 1000

Ground-truth semantic IDs are Waymo class IDs. During evaluation, they are mapped to the release train IDs listed below. Raw semantic ID 0 (undefined) and raw semantic ID 5 (motorcyclist) are ignored.

For stuff classes, all points of the same semantic class are treated as a single segment during evaluation.

Classes

Waymo ID	Train ID	Name	Type
0	0	undefined	void
1	1	car	thing
2	2	truck	thing
3	3	bus	thing
4	4	other vehicle	thing
5	0	motorcyclist	void
6	5	bicyclist	thing
7	6	pedestrian	thing
12	7	bicycle	stuff
13	8	motorcycle	stuff
8	9	sign	stuff
9	10	traffic light	stuff
10	11	pole	stuff
11	12	construction cone	stuff
14	13	building	stuff
15	14	vegetation	stuff
16	15	tree trunk	stuff
17	16	curb	stuff
18	17	road	stuff
19	18	lane marker	stuff
20	19	other ground	stuff
21	20	walkable	stuff
22	21	sidewalk	stuff

The evaluation classes are train IDs 1..21; train ID 0 is ignored.

The class mapping is also provided in panoptic_waymo_label_mapping.yaml.

Evaluation

Evaluation is performed using the Panoptic Waymo Devkit.

Prediction Format

Predictions must be stored as one .npz file per validation frame in a dedicated directory:

predictions/
  <frame_id>_<segment_id>.npz

The filename must match the manifest lidar_token. For example:

000025_segment-10203656353524179475_7625_000_7645_000.npz

Each prediction file must contain key data, have shape (num_points,), and be convertible to uint32.

Predictions are encoded using train IDs:

pred_panoptic = train_id * 1000 + pred_instance_id

Encoding Rules:

Predictions are encoded using a combination of the semantic train ID and the instance ID:

pred_panoptic = train_id * 1000 + pred_instance_id

• Void/Ignore: Train ID 0 is ignored.
• Valid Classes: Train IDs are 1..21.
• Things: Instance IDs for thing classes must be positive, per-frame unique instance IDs.
• Stuff: Instance IDs for stuff classes are ignored by the evaluator and will be collapsed to train_id * 1000
• Prediction files are required for every frame in manifests/val.jsonl.

Reported Metrics

• PQ: mean Panoptic Quality
• SQ: mean Segmentation Quality
• RQ: mean Recognition Quality
• mIoU: mean semantic intersection-over-union
• PQ_dagger: PQ for thing classes and IoU for stuff classes
• thing-class and stuff-class metrics
• per-class PQ, SQ, RQ, and IoU

Please refer to the Panoptic Waymo devkit repository for installation, visualization, and evaluation commands.

Citation

If you use Panoptic Waymo or the devkit in your research, please cite our paper:

UP-Fuse: Uncertainty-guided LiDAR-Camera Fusion for 3D Panoptic Segmentation (Robotics: Science and Systems, 2026)

@article{mohan2026upfuse,
  title={UP-Fuse: Uncertainty-guided LiDAR-Camera Fusion for 3D Panoptic Segmentation},
  author={Mohan, Rohit and Drews, Florian and Miron, Yakov and Cattaneo, Daniele and Valada, Abhinav},
  journal={arXiv preprint arXiv:2602.19349},
  year={2026}
}

Contact

For technical questions, bug reports, or evaluation issues related to the Panoptic Waymo benchmark or devkit, please open an issue in the Panoptic Waymo Devkit repository.

For other inquiries, academic collaborations, or specific questions, please contact Rohit Mohan at mohan@cs.uni-freiburg.de.