Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): ('json', {}), NamedSplit('test'): ('webdataset', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

SO-RoadUAV Dataset (Benchmark-V3)

SO-RoadUAV is a CARLA-first benchmark for map-assisted low-altitude road inspection from single-frame UAV RGB images. This directory is a deterministic public serialisation of the frozen Benchmark-V3-1.0 release.

Status: released.

Naming. The benchmark is referred to as Benchmark V3 (and benchmark_v3 in code identifiers and historical artifacts) throughout its development. Its public name is SO-RoadUAV; both names denote the same dataset and the same protocol version.

License. The dataset is released under CC BY-NC 4.0 (Attribution-NonCommercial 4.0 International): reuse and adaptation are permitted with attribution, commercial use is not. See LICENSE.

Third-party assets. The RGB images are renderings of CARLA map assets, distributed by the CARLA project under CC-BY (CVC/UAB). The CC BY-NC 4.0 license applies to the SO-RoadUAV curation, annotations and packaging; it does not restrict the CARLA map assets themselves.

Distribution

The package is published as four tar archives under archives/, because the Hub limits a Git-backed directory to 10,000 entries and train/rgb, train/metadata and train/lane_polyline_projected each contain 12,000 files:

Archive Contents Size
SO-RoadUAV-1.0-train.tar train/ 3.6 GB
SO-RoadUAV-1.0-val.tar val/ 0.5 GB
SO-RoadUAV-1.0-test.tar test/ 1.4 GB
SO-RoadUAV-1.0-test-eval.tar test/road_mask_visible, test/lane_polyline_projected, test/metadata 0.24 GB

The archives preserve the directory layout below; extracting them into one directory reproduces this package exactly. archives/SHA256SUMS covers the archives, and SHA256SUMS covers the 74,508 files of the extracted package.

python scripts/download_dataset.py --dest dataset downloads, checks and extracts the archives automatically. Manual extraction:

mkdir -p dataset
tar -xf SO-RoadUAV-1.0-train.tar -C dataset
tar -xf SO-RoadUAV-1.0-val.tar   -C dataset
tar -xf SO-RoadUAV-1.0-test.tar  -C dataset

The top-level text files (README.md, LICENSE, SHA256SUMS, manifests/, the two JSON cards) are not inside the archives; download them alongside the archives when assembling the package by hand.

Tasks and labels

  1. Road-region segmentation — target road_mask_visible, a binary PNG (0 / 255). Foreground is the CARLA semantic mask for Roads=1 plus RoadLines=24 in the rendered frame.
  2. Target-corridor lane recovery — target lane_polyline_projected, a JSON document whose instances[*].points_image lists image-space lane-marking polylines with projected_in_image_no_occlusion_test semantics. Each instance is one distinct physical lane marking. The evaluation ROI is target_corridor_polygon from the sample metadata; predictions outside that ROI are not target-corridor evidence.

No paired Visible/Full targets are published. Road heading, corridor geometry, lane graphs, topology tags and map priors are provided for slicing and QA only.

Layout

train/  val/  test/
    rgb/                     JPEG, 1280 x 720
    road_mask_visible/       PNG, 0 / 255
    lane_polyline_projected/ JSON, at most 10 instances per image
    metadata/                cleaned per-sample metadata
manifests/
    subsets/{train,val,test}.jsonl
    views/{core_scene,core_observation,core_axis,challenge_l18}.jsonl
dataset_card.json
package_manifest.json
LICENSE
SHA256SUMS

Every sample appears exactly once under train/, val/ or test/. The four views/ manifests are logical indexes over the same test/ assets; they do not duplicate images.

Use the manifests as the entry point. Do not construct the dataset by scanning directories: the manifests carry the identity fields, subset routing and slice membership that the evaluation protocol depends on.

Each subset manifest row resolves four relative asset paths (rgb, road_mask, lane, metadata) against this directory, and repeats the condition summary, reference linkage, factor labels and label_rule_version.

Logical views

All four views are subsets of test/.

View Images Meaning
core_scene 5,040 every test image, for scene-level aggregation
core_observation 1,890 one-factor-at-a-time slice around each anchor reference
core_axis 882 seven stress axes
challenge_l18 2,268 multifactor challenge rows

Composition

Subset Images Sites Anchors
train 12,000 131 250
val 1,584 20 33
test 5,040 53 126

The test split covers 53 sites, 126 anchors and 25 leakage groups, with 40 unique images per anchor (41 logical slots including the reference alias). Scenes: straight 1–4 and 5–8 lanes, curve 1–4 and 5–8 lanes, intersection and ramp.

Observation ranges: height 20–100 m, pitch −90°…−70°, yaw −15°…+15°, traffic 0–40 vehicles/lane/km, three weather families (dry, rain-wet, fog) and sun altitude −5°…75°.

Label rules

Two label-rule versions are published:

label_rule_version Samples Rule
lane_marking_projection_default_1.0 15,800 OpenDRIVE marking projection with per-span de-duplication
lane_marking_physical_merge_1.0 2,824 physical cross-span marking merge at 24 confirmed sites

metadata.json records which rule produced each sample. Lane instances are already final: do not merge or split them downstream.

Integrity

SHA256SUMS lists every payload file plus the two READMEs, both JSON cards, the LICENSE and the seven manifests. Verify from this directory:

sha256sum -c SHA256SUMS

Citation and license

License: CC BY-NC 4.0 — see LICENSE. Reuse and adaptation are permitted with attribution; commercial use is not.

Citation:

Wang, Z., Zhao, J., Liu, X., Gao, X., & Ye, K. SO-RoadUAV: A Scene-Observation Diagnostic Benchmark for UAV Traffic-Monitoring Road Perception (2026).

The images are rendered with the CARLA simulator from its map assets; consult the CARLA project for the terms applicable to those assets.

Downloads last month
313