README.md

Roadwork Cones Dataset

This dataset is designed for detecting roadwork-zone objects in autonomous driving scenarios. It contains three classes - traffic cones, roadworks signs, and vertical guide panels (delineators) - captured from four vehicle-mounted cameras across 39 driving sessions in urban and suburban roads.

Dataset Description

• Source: Fleet of autonomous vehicles, urban and suburban roads
• Sensor: 4x LUCID TRI054S-CC cameras (2880×1860), front/side facing
• Coverage: ~31K bounding box annotations, ~4.7K unique images, 3 classes
• Splits: Train 70% (22,841 rows, 2,871 images), Test 30% (8,561 rows, 1,803 images) - split by driving session to prevent temporal leakage
• Format: Parquet (annotations) + JPEG (images) in train/ and test/ subdirectories

Classes (3)

Class	Count	Description
cone	11,520	Standard traffic cone
roadworks	2,813	Roadwork zone sign / panel
vertical_pannel	17,069	Vertical guide panel (delineator)

Data Fields

Field	Type	Description
bbox_msg_id	VARCHAR	UUID linking to the original bounding box message
object_id	BIGINT	Unique object tracking ID
label	VARCHAR	Object class (cone, roadworks, vertical_pannel)
bbox_coords	DOUBLE[4]	Bounding box [x, y, width, height] in pixel coordinates
timestamp_ns	BIGINT	ROS bag timestamp (nanoseconds)
image_path	VARCHAR	Relative path to JPEG in train/ or test/

Data Splits

Split	Rows	Images
train	22,841	2,871
test	8,561	1,803

Dataset Structure

roadwork_cones_dataset/
├── annotations.parquet   # All annotations (31,402 rows)
├── train/
│   ├── camera_1C0FAF5250E2/   # 854 images
│   ├── camera_1C0FAF57D6F8/   # 1,415 images
│   ├── camera_1C0FAF5CA7B6/   # 494 images
│   └── camera_1C0FAF5CC14D/   # 108 images
└── test/
    ├── camera_1C0FAF5250E2/   # 570 images
    ├── camera_1C0FAF57D6F8/   # 1,193 images
    ├── camera_1C0FAF5CA7B6/   # 40 images
    └── camera_1C0FAF5CC14D/   # 0 images

Usage

import pandas as pd

df = pd.read_parquet("annotations.parquet")
print(f"{len(df)} annotations, {df.image_path.nunique()} unique images")