PanoHK360

A Large-Scale 8K Urban Panoramic RGB-D Dataset for Depth Estimation and 3D Perception

8K equirectangular RGB panoramas with LiDAR-derived metric depth, surface normals, point clouds, and 6-DoF poses from continuous outdoor driving sequences.

Overview · Download · Data Contents · Evaluation · Limitations · Citation

---

Dataset Teaser

Open the RGB–LiDAR alignment video

Temporal RGB–LiDAR correspondence along an urban Hong Kong trajectory.

The teaser illustrates how continuously captured panoramic frames are aligned with synchronized LiDAR measurements to generate metric geometric annotations.

---

Highlights

LiDAR-grounded metric depth Valid depth observations are obtained from calibrated LiDAR measurements projected into the panoramic image domain. They are not generated by a monocular teacher model or synthetic renderer.	Native 8K panoramic imagery RGB frames are provided in an 8000 × 4000 equirectangular projection, preserving wide-field urban context and fine-grained scene detail.
Multi-modal geometric annotations The release combines RGB panoramas, metric depth, surface normals, LiDAR point clouds, and 6-DoF poses for geometry-aware learning and reconstruction.	Continuous outdoor sequences Vehicle-mounted acquisition captures street canyons, intersections, façades, vegetation, traffic participants, and challenging occlusion patterns.

---

Dataset Overview

Property	Description
Dataset name	PanoHK360
Domain	Outdoor urban scenes
Location	Hong Kong
Image projection	Equirectangular projection (ERP)
RGB resolution	8000 × 4000 pixels
Modalities	RGB panorama, metric depth, surface normals, LiDAR point cloud, and 6-DoF pose
Depth source	Survey-grade LiDAR
Panoramic camera	Teledyne FLIR Ladybug
LiDAR scanner	RIEGL VUXR-1HA22
Current repository release	R101 20230413--filter
Compressed release size	Approximately 2.77 GB
License	CC BY 4.0

Supported research tasks

PanoHK360 is intended to support research in:

• panoramic monocular depth estimation;
• RGB–LiDAR fusion and registration;
• 360° geometric scene understanding;
• surface-normal estimation;
• point-cloud generation and completion;
• temporal geometric consistency;
• camera localization and trajectory-aware reconstruction;
• urban 3D reconstruction;
• robustness analysis for ERP-aware vision models.

---

Sensor Platform

The dataset was collected using a vehicle-mounted, time-synchronized sensor platform that combines panoramic RGB imaging with survey-grade LiDAR scanning. Calibration and synchronization enable LiDAR measurements to be transformed into the panoramic camera coordinate system.

Teledyne FLIR Ladybug 360° panoramic RGB imaging

RIEGL VUXR-1HA22 Survey-grade LiDAR scanning

---

Data Acquisition and Annotation

The released data are produced through the following high-level pipeline:

1. Panoramic RGB frames and LiDAR sweeps are captured along continuous driving trajectories.
2. Sensor streams are time-synchronized and transformed into a shared calibrated coordinate system.
3. LiDAR returns are projected into the equirectangular image domain to obtain metric depth observations.
4. Frame-aligned geometric products, including surface normals, point clouds, and 6-DoF poses, are stored alongside the RGB panoramas.
5. A filtered subset is packaged as the current public repository release.

Because LiDAR samples visible surfaces discretely, the depth maps may contain invalid or unobserved pixels. Users should preserve the provided validity convention when constructing training targets and evaluation masks.

---

Repository Structure

PanoHK360/
├── README.md
├── R101 20230413--filter/          # Browsable filtered release
├── R101 20230413--filter.zip       # Compressed copy of the same release (~2.77 GB)
├── show.mp4                        # RGB–LiDAR alignment teaser
├── ladybug_image.png               # Panoramic camera reference image
├── RIEGL VUXR-1HA22.png            # LiDAR scanner reference image
└── .gitattributes

R101 20230413--filter/ and R101 20230413--filter.zip contain the same filtered release. Browse the directory on the Hub or download the archive for local experiments.

---

Download

Replace <namespace>/PanoHK360 with the repository ID displayed at the top of your Hugging Face dataset page.

Hugging Face CLI

pip install -U huggingface_hub

hf download <namespace>/PanoHK360 \
  "R101 20230413--filter.zip" \
  --repo-type dataset \
  --local-dir ./PanoHK360

unzip "./PanoHK360/R101 20230413--filter.zip" \
  -d ./PanoHK360/release

Python

from pathlib import Path
from zipfile import ZipFile

from huggingface_hub import hf_hub_download

repo_id = "<namespace>/PanoHK360"
local_dir = Path("PanoHK360")
local_dir.mkdir(parents=True, exist_ok=True)

archive_path = hf_hub_download(
    repo_id=repo_id,
    filename="R101 20230413--filter.zip",
    repo_type="dataset",
    local_dir=local_dir,
)

extract_dir = local_dir / "release"
with ZipFile(archive_path, "r") as archive:
    archive.extractall(extract_dir)

print(f"Dataset extracted to: {extract_dir.resolve()}")

Browse on the Hub

The extracted release can also be inspected directly without downloading the archive:

R101 20230413--filter/

Large files on the Hugging Face Hub may be backed by Git LFS or Xet. Use an up-to-date huggingface_hub installation when downloading the release.

---

Data Contents

The current release provides frame-level assets for panoramic appearance and geometry.

Modality	Description	Example use
RGB panorama	High-resolution equirectangular color image	Model input and panoramic perception
Metric depth	LiDAR-derived depth aligned with the panorama	Supervised depth learning and evaluation
Surface normals	Geometric surface orientation aligned with each frame	Normal estimation and geometry-aware learning
LiDAR point cloud	3D measurements associated with the captured scene	Registration, fusion, and reconstruction
6-DoF pose	Frame or platform pose information	Temporal alignment and trajectory reconstruction

Cross-modal correspondence

When implementing a data loader:

1. preserve the original frame identifiers;
2. match modalities using the provided naming convention;
3. retain invalid-depth masks or sentinel values;
4. verify coordinate-system conventions before transforming point clouds or poses;
5. record any resizing, cropping, interpolation, or depth densification applied during preprocessing.

The repository distributes the original frame-level files rather than a standardized Hugging Face datasets table. Consequently, the exact loader implementation should follow the directory and naming conventions present in the downloaded release.

---

Minimal Inspection Script

The following script lists the extracted files without assuming undocumented file extensions or subdirectory names:

from collections import Counter
from pathlib import Path

root = Path("PanoHK360/release")

if not root.exists():
    raise FileNotFoundError(
        f"Dataset directory not found: {root.resolve()}"
    )

files = [path for path in root.rglob("*") if path.is_file()]
extension_counts = Counter(path.suffix.lower() or "<no extension>" for path in files)

print(f"Root: {root.resolve()}")
print(f"Total files: {len(files):,}")
print("Extensions:")
for extension, count in extension_counts.most_common():
    print(f"  {extension}: {count:,}")

print("\nFirst 20 files:")
for path in files[:20]:
    print(path.relative_to(root))

---

Recommended Evaluation Practices

To make comparisons reproducible, report the following details with experimental results:

• the dataset release or Hub revision used;
• the train, validation, and test split definition;
• whether geographically or temporally adjacent frames were separated across splits;
• the input resolution and whether the full ERP image, crops, or cube-map projections were used;
• the valid-depth mask and evaluated depth range;
• whether predictions were evaluated in metric scale or after scale alignment;
• interpolation and resizing methods for RGB, depth, and normal maps;
• handling of sparse LiDAR observations and moving objects;
• the exact depth metrics and spherical weighting strategy, when applicable.

For ERP evaluation, users should consider latitude-dependent pixel area. Unweighted image-space metrics may overrepresent regions near the poles of the equirectangular panorama.

---

Limitations and Responsible Use

Geographic and environmental scope

PanoHK360 represents selected outdoor routes in Hong Kong. It should not be assumed to cover all cities, road types, architectural styles, seasons, weather conditions, illumination conditions, or traffic patterns.

Viewpoint bias

Vehicle-mounted acquisition emphasizes road-level observations. Pedestrian-only areas, indoor scenes, elevated viewpoints, narrow passages, and inaccessible locations may be underrepresented.

Equirectangular distortion

ERP imagery exhibits latitude-dependent distortion, particularly near the top and bottom of the panorama. Planar convolution, resizing, and image-space evaluation can introduce geometric bias unless spherical structure is considered.

Sensor and annotation limitations

LiDAR-derived annotations may be affected by occlusion, limited sampling density, range limits, calibration residuals, synchronization errors, reflective or transparent materials, and independently moving objects.

Privacy

Urban imagery may include people, faces, vehicle license plates, storefronts, residences, or other identifiable content. Users are responsible for reviewing the data and complying with applicable privacy, data-protection, and local legal requirements before redistribution or deployment.

Safety-critical deployment

This dataset is released for research. It should not be used as the sole basis for autonomous-driving, surveillance, mapping, or other safety-critical decisions without independent validation, risk assessment, and appropriate human oversight.

---

Citation

If you use PanoHK360 in academic work, please cite the associated paper after its final bibliographic information becomes available. Until then, use the provisional entry below:

@misc{panohk360,
  title        = {PanoHK360: A Large-Scale 8K Urban Panoramic Dataset and Benchmark for Depth Estimation},
  author       = {Anonymous Authors},
  year         = {2026},
  note         = {Under review}
}

---

License

PanoHK360 is released under the Creative Commons Attribution 4.0 International License.

When using or redistributing the dataset, figures, or derived annotations, provide appropriate attribution and clearly describe any modifications.

---

Contact and Issues

For questions about data organization, annotations, or benchmark usage, please open an issue in the Hugging Face dataset repository. When reporting a problem, include the affected frame identifier, file path, and repository revision whenever possible.

_{Built for panoramic depth estimation, RGB–LiDAR fusion, and urban 3D perception research.}