WideDrive

WideDrive is a CARLA synthetic dataset for wide-field-of-view novel-view synthesis: three rear-facing source cameras A + B + D → C_RECT, with synchronized LiDAR and camera/vehicle poses.

Snapshot status — 2026-09-26

This folder contains 120 structurally validated clips, split into 100 training clips and 20 test clips. Each clip contains 200 synchronized frames captured at 10 Hz over a nominal 20-second window. The first-to-last frame timestamp span is about 19.9 seconds.

The requested dynamic-scene refresh is incomplete. Of 34 clips selected for replacement, 5 replacements passed and 29 remain below the motion gate. Across the complete snapshot, 91 clips meet that gate. The remaining 29 are included in their original, structurally complete form; this snapshot must not be described as a fully motion-validated release. See reports/dataset_audit.json and reports/motion_pending.txt.

The motion gate requires at least 20 m of horizontal ego travel and at least 50% of the capture duration moving faster than 0.5 m/s. Motion is measured from all saved ego poses; image noise is not counted as camera movement.

ZIP distribution

The scene payloads in this transfer folder are distributed as one ZIP archive per scene, for example Town01_scene_0001.zip. Each ZIP contains its own Town01_scene_0001/ top-level directory; relative payload paths are preserved. The folder structure documented below is the structure after extraction.

Extract one scene from this folder with:

unzip Town01_scene_0001.zip -d .

Extract all downloaded scenes with:

for archive in Town*_scene_*.zip; do
    unzip "$archive" -d .
done

The archives use ZIP64 where required, lossless deflate for point clouds and metadata, and store the already-compressed PNGs. Each published .zip has passed a complete entry-size comparison and CRC check. Hidden .Town*.partial.zip files are still being created and are not upload artifacts.

archives_manifest.json records each completed archive's byte size, split and SHA-256. Packaging is finished only when its state is complete and complete is 120. To verify the downloaded archives:

sha256sum -c SHA256SUMS

Dataset overview

Property	Value
Simulator / capture quality	CARLA 0.9.16 / Epic capture protocol
Distinct town layouts	Town01, Town03, Town04, Town05, Town10HD
Actual maps	Town01, Town03_Opt, Town04, Town05_Opt, Town10HD
Clips per town	24: 20 train + 4 test
Weather	ClearNoon and ClearNight; 60 clips each
Frames	200 per clip; 24,000 synchronized frames in total
RGB images	4 channels × 200 per clip; 96,000 PNGs in total
LiDAR scans	200 per clip; 24,000 .pcd.bin scans in total
LiDAR returns	517,491–698,370 points per frame
Keyframes	2 Hz: 40 per clip, 4,800 in total
NPCs at capture setup	5 vehicles and 3 walkers per clip

_Opt denotes a layered variant of the corresponding town, not a new town layout. Day/night capture and the train/test labels are recorded in each clip's metadata/config.json.

Folder structure

WideDrive/
├── README.md
├── dataset_index.json
├── render_plan.json                    # Original capture plan; see actual config for seeds
├── splits/
│   ├── train.txt                       # 100 clip names, one per line
│   └── test.txt                        # 20 clip names, one per line
├── reports/
│   ├── dataset_audit.json               # Current structural and motion audit
│   └── motion_pending.txt              # 29 clips still needing dynamic replacement
├── docs/
│   ├── SCHEMA.md
│   └── COORDINATES.md
└── TownXX_scene_NNNN/                   # 120 scene folders directly under this root
    ├── samples/<CHANNEL>/               # 40 keyframe payloads per channel
    ├── sweeps/<CHANNEL>/                # 160 other payloads per channel
    ├── v1.0-trainval/
    │   ├── sensor.json
    │   ├── calibrated_sensor.json
    │   ├── ego_pose.json
    │   ├── sample.json
    │   ├── sample_data.json
    │   ├── scene.json
    │   └── log.json
    ├── metadata/
    │   ├── frames.json                 # Recommended frame-level entry point
    │   ├── config.json
    │   ├── rig.json
    │   ├── coordinate_systems.json
    │   ├── task_definition.json
    │   └── splits.json
    └── debug/                          # Capture diagnostics and validation evidence

Some clips additionally contain motion.json, render_provenance.json or delivery.json. Their absence in earlier clips does not imply missing sensor frames. Batch logs, failed temporary captures and old-scene backups are outside this delivery folder.

Sensor channels

Manifest key	Payload channel	Resolution / format	Role
source.A	CAM_BACK_LEFT_SOURCE	1920 × 1080 RGB PNG; 60° horizontal FoV	Input
source.B	CAM_BACK_SOURCE	1920 × 1080 RGB PNG; 60° horizontal FoV	Input
source.D	CAM_BACK_RIGHT_SOURCE	1920 × 1080 RGB PNG; 60° horizontal FoV	Input
ground_truth.C_rect	CAM_BACK_WIDE_RECT_GT	5760 × 1080 RGB PNG; 120° horizontal FoV	Withheld ground truth
lidar	LIDAR_TOP	32-channel, 360° LiDAR; little-endian float32	Geometry

C_RECT shares B's camera pose. A/B/D have distinct physical camera centers. C_RECT is never listed as a source camera; follow each clip's task_definition.json for the source/GT separation. No cylindrical GT channel, masks or duplicate native LiDAR arrays are included.

LiDAR files have no header and contain N × 5 float32 values in column order [x, y, z, intensity, ring]. XYZ is in metres. The ring is stored as a float32; check lidar_metadata.ring_valid before using it as an integer. A point takes 20 bytes, so the point count is file_size_bytes / 20.

Splits and naming

Each town has scene_0001 through scene_0024. Numbers 0001–0020 are training clips and 0021–0024 are test clips. Odd numbers use ClearNoon; even numbers use ClearNight. Read the split lists or dataset_index.json rather than relying on filesystem ordering.

The split separates clip names; the same town layouts appear in both splits. It is not a held-out-town evaluation split. There is no validation split.

Read one synchronized frame

Reading the dataset does not require a running CARLA server. With NumPy and Pillow installed, run this example from the dataset folder:

import json
from pathlib import Path

import numpy as np
from PIL import Image

root = Path(".")
scene_name = (root / "splits/train.txt").read_text().splitlines()[0]
scene = root / scene_name
frames = json.loads((scene / "metadata/frames.json").read_text())
frame = frames[0]

sources = {name: np.array(Image.open(scene / path))
           for name, path in frame["source"].items()}
gt = np.array(Image.open(scene / frame["ground_truth"]["C_rect"]))
lidar = np.fromfile(scene / frame["lidar"], dtype="<f4").reshape(-1, 5)

assert len(frames) == 200
assert lidar.shape[0] == frame["lidar_metadata"]["point_count"]
print(sources["B"].shape, gt.shape, lidar.shape)

Payload paths in frames.json and sample_data.json are relative to their scene folder, not the dataset root. Resolve them with scene / path. Historical absolute paths in capture configs/provenance describe where the capture ran; they are not needed to read this relocated dataset.

Coordinates and timestamps

• Global, ego and LiDAR coordinates are right-handed: +X forward, +Y left, +Z up. Camera coordinates follow OpenCV: +X right, +Y down, +Z forward.
• Quaternions use [w, x, y, z]; extrinsics map sensor coordinates into ego coordinates. The ego reference is based on the rear axle, with the method recorded in rig.json.
• timestamp_us is simulation time in microseconds. Frame intervals are 100,000 µs, allowing a 1 µs rounding difference. These are not wall-clock acquisition dates, and different clips may have overlapping timestamps.
• Frame records include sensor/ego/global transform matrices. See docs/COORDINATES.md for conversion details.

nuScenes-style metadata and validation

This is nuScenes-like data, with seven tables stored separately for each clip, five custom sensor channels, and 1,000 sample_data rows per clip. Non-keyframe records point to the latest preceding 2 Hz keyframe. Channel prev/next links follow the 10 Hz capture sequence.

It is not a complete official nuScenes release: no object/category/3D-box annotation tables are provided, and the per-clip tables cannot be passed to an unmodified nuScenes SDK loader as one global dataset. For NVS, use metadata/frames.json. See docs/SCHEMA.md.

The packaging audit checks all 120 clips for table references, payload existence and image dimensions, frame/timestamp associations, per-channel chains, point counts and LiDAR file sizes, camera poses, traffic spawn counts, split/weather configuration and absence of unwanted channels. It also measures motion from every saved ego pose. It does not fully decode every PNG pixel or rerun the simulator. Per-clip capture diagnostics remain in debug/.

No dataset license is specified in this snapshot.