Dur360BEV: A Real-world 360-degree Single Camera Dataset and Benchmark for Bird-Eye View Mapping in Autonomous Driving

Abstract

We present Dur360BEV, a novel spherical camera autonomous driving dataset featuring a 128-channel 3D LiDAR and an RTK-refined GNSS/INS system, along with a benchmark for Bird’s Eye View (BEV) map generation using a single spherical camera. This dataset addresses the challenges of BEV generation by reducing hardware complexity, replacing multiple perspective cameras with a single 360-degree camera.

Our benchmark includes SI2BEV (Spherical-Image-to-BEV), a novel module that projects spherical imagery into BEV space with a refined sampling strategy, as well as an adaptive Focal Loss formulation to handle extreme class imbalance in BEV segmentation. Extensive experiments demonstrate that Dur360BEV simplifies the sensor setup while achieving competitive performance.

News

• [2025/01/27] Dur360BEV has been accepted by ICRA 2025.

Sensor placement

Dur360BEV is equipped with a spherical camera, 3D LiDAR, and a high-precision GNSS/INS system, providing comprehensive environmental perception.

• LiDAR: Ouster OS1-128 LiDAR sensor with 128 channels vertical resolution
• Spherical Camera: Ricoh Theta S featuring a dual 1/2.3" 12M CMOS sensor, 1280 × 640 resolution, and 15 Hz capture frequency
• GNSS/INS: OxTS RT3000v3 global navigation satellite and inertial navigation system, supporting localization from GPS, GLONASS, BeiDou, Galileo, PPP and SBAS constellations

Dataset File Description

dataset/ 
├── image
│   ├── data
│   └── timestamps.txt
├── labels
│   ├── data
│   ├── dataformat.txt
│   └── timestamps.txt
├── md5sums.txt
├── metadata
│   └── os1.json
├── ouster_points
│   ├── data
│   ├── dataformat.txt
│   └── timestamps.txt
└── oxts
    ├── data
    ├── dataformat.txt
    └── timestamps.txt

Dataset Download

Links

Train and validation split used in our paper

• Download from Hugging Face

Dur360BEV-Extended

• Download from Hugging Face

3D model used in our LiDAR-Camera Setup

To do...

Download Guidance

Downloading the full dataset from Hugging Face

Prerequisites

• Before starting, ensure you have the Hugging Face CLI (huggingface_hub) installed:

  ​ Install it via: pip install -U "huggingface_hub[cli]"

• Since the Dur360BEV dataset is a gated (restricted access) dataset on Hugging Face, you need to authenticate before downloading it.

  • You first need a Hugging Face account. If you don’t have one, please register.
  • Authenticate via the command line on the computer where you want to download the dataset by entering: huggingface-cli login. Following the instructions of that command, and it will prompt you for your Hugging Face Access Token.
  • Open this link and login to your Hugging Face account. At the top of the page, in the section “You need to agree to share your contact information to access this dataset”, agree to the conditions and access the dataset content. If you have already agreed and been automatically granted access, the page will display: “Gated dataset: You have been granted access to this dataset.”

Download the dataset using scripts

We provide a pre-written Bash script to download the dataset from Hugging Face. You need to manually modify the User Configuration (Modify as Needed) section at the beginning of dur360bev_hf_download_script.sh to match your desired paths and features.

If you encounter any issues (e.g., network problems or unexpected interruptions), you can also modify this script to fit your needs. For example, if the extraction process is interrupted, you can manually comment out the dataset download section and resume from extraction.

Note: The Hugging Face service may be restricted in certain countries or regions. In such cases, you may consider using a Hugging Face mirror (Hugging Face 镜像) as an alternative.

Environment Setup

To ensure reproducibility and ease of installation, we provide a structured virtual environment setup. This includes all required dependencies and the local fisheye_tools library.

git clone https://github.com/yourusername/icra2025-dur360bev.git
cd icra2025-dur360bev

Virtual Envrionment setup

python3 -m venv dur360bev
source dur360bev/bin/activate
pip install -r requirements.txt