PanoCARLA-Static

PanoCARLA-Static is a variant of PanoCARLA captured without dynamic objects, designed for tasks that require static scenes while the camera roams through the environment. Generated with CARLA 0.9.16, it provides synchronized panoramic RGB, metric depth, semantic labels, and camera poses along routes through eight towns.

Here, static refers to the scene, not the camera: dynamic vehicle and pedestrian actors are removed, while the camera continues to move through the environment with varying height. The dataset supports studying panoramic scene geometry and temporal consistency during static-scene roaming.

Relationship to PanoCARLA

PanoCARLA-Static follows the same overall roaming routes as PanoCARLA, but the full 3D camera trajectories are not identical. The horizontal movement follows the original routes, while the camera-height variation is generated again using a new random realization for the static capture.

Corresponding route names or source frame IDs therefore do not imply identical camera poses or views. PanoCARLA-Static is not a strictly pose-matched, frame-by-frame pair of the original dataset with moving objects removed. Use the poses supplied with each release for geometric processing or cross-version comparisons. Collision filtering and the resulting clip boundaries can also differ between the two releases.

Data Collection Code

The data collection pipeline used to construct PanoCARLA and PanoCARLA-Static is publicly available at PanoCARLA GitHub Repository. It includes trajectory recording, panoramic data capture, and panorama stitching scripts.

Dataset Overview

The released HDF5 data contains 127,334 frames in 127 clips from 67 routes across 8 towns.

Per-town Statistics

Town	Routes	Clips	Frames
Town01	8	23	15,613
Town02	6	14	6,856
Town03	8	23	17,695
Town04	8	16	23,701
Town05	10	19	25,429
Town06	8	10	15,784
Town07	12	13	15,012
Town10	7	9	7,244

Capture runs at 20 FPS. Frames marked in the collision logs are excluded, sequences are split at collisions or gaps in frame IDs, and continuous segments shorter than 100 frames are discarded.

Repository Structure

PanoCARLA-Static/
├── panocarla_static_h5/          # HDF5 clips (2400 x 1200)
├── setting_static.json          # Hierarchical clip/frame index with HDF5 references
├── statistics_report_static.txt # Detailed per-town and per-clip statistics
├── carla_path_vis/              # Route visualizations shared with PanoCARLA
├── vis_mp4/                     # Full-route RGB/depth preview videos
└── README.md

All released frames are stored in panocarla_static_h5/.

The eight route images in carla_path_vis/ are the same as those in PanoCARLA, since both datasets share the overall roaming routes.

vis_mp4/ contains 67 previews from the static capture. Each video runs at 20 FPS and stacks a 1024 x 512 RGB panorama above its depth visualization, producing a 1024 x 1024 frame. These previews cover the full captured routes, including frames subsequently excluded from HDF5, so video frame indices are not HDF5 indices. The HDF5 panoramas retain the higher 2400 x 1200 resolution.

HDF5 Format

One HDF5 file holds one continuous clip; T denotes its frame count.

Key	Shape	dtype	Description
rgb	[T, 1200, 2400, 3]	uint8	Panoramic color images in RGB order
depth	[T, 1200, 2400, 1]	float16	Radial distance from the camera, in meters
seg	[T, 1200, 2400, 1]	uint8	Per-pixel semantic class IDs
poses	[T, 4]	float32	Camera position and heading: x, y, z, yaw
frame_ids	[T]	int32	Frame identifiers from the source capture

Positions use CARLA world coordinates in meters; yaw is in degrees. The camera has zero pitch and roll, so these components are omitted. Semantic labels follow the CARLA 0.9.16 specification.

Reading an HDF5 Clip

import h5py

h5_path = "PanoCARLA-Static/panocarla_static_h5/town01_path1_clip_0.h5"
with h5py.File(h5_path, "r") as clip:
    rgb = clip["rgb"][0]       # [1200, 2400, 3]
    depth = clip["depth"][0]   # [1200, 2400, 1], meters
    seg = clip["seg"][0]       # [1200, 2400, 1]
    pose = clip["poses"][0]    # x, y, z, yaw
    frame_id = int(clip["frame_ids"][0])

Frame Index

setting_static.json retains the original index hierarchy:

town -> path -> clip -> list of frame records

Each record contains frame_id, h5_path, h5_index, and pose, whose keys are x, y, z, and yaw. The HDF5 path is relative to the repository root, and the index is zero-based within that file. Pose values are taken from the finalized HDF5 data.

Unlike the original raw-data index, this JSON directly references the published HDF5 files rather than PNG/NPY paths:

import json
from pathlib import Path
import h5py

root = Path("PanoCARLA-Static")
index = json.loads((root / "setting_static.json").read_text())
frame = index["town01"]["path1"]["town01_path1_clip_0"][0]

with h5py.File(root / frame["h5_path"], "r") as clip:
    i = frame["h5_index"]
    rgb = clip["rgb"][i]
    pose = clip["poses"][i]
    assert int(clip["frame_ids"][i]) == frame["frame_id"]

Download

Download with the Hugging Face CLI. If access approval is enabled on the repository, request access on the dataset page and authenticate first:

hf auth login

Download all HDF5 clips and their frame index:

hf download Soon122/PanoCARLA-Static \
  --repo-type dataset \
  --include "panocarla_static_h5/*" \
  --include "setting_static.json" \
  --include "statistics_report_static.txt" \
  --local-dir PanoCARLA-Static

Download the preview videos and route images:

hf download Soon122/PanoCARLA-Static \
  --repo-type dataset \
  --include "vis_mp4/*" \
  --include "carla_path_vis/*" \
  --local-dir PanoCARLA-Static

License

PanoCARLA-Static follows PanoCARLA's CC BY 4.0 license. Please provide attribution when using or redistributing the dataset.

Citation

Please cite the PVDepth paper:

@inproceedings{song2026pvdepth,
  author    = {Song, Chuanxin and Peng, Peixi},
  title     = {PVDepth: Panoramic Video Depth Estimation via Geometry-Aware Spatiotemporal Adaptation},
  booktitle = {ICML},
  year      = {2026}
}

Acknowledgements

This dataset was captured with CARLA 0.9.16. We thank the CARLA team for the simulator and its environments.