BJTU-UVA: The First Dataset for Automatic Spectral Calibration of Hyperspectral Images

We are proud to introduce BJTU-UVA, the first dataset designed specifically for the task of automatic spectral calibration of hyperspectral images (HSIs). This dataset addresses the critical challenge of minimizing illumination variability without relying on manual intervention or physical references.

Key Highlights

• Task Proposal:
  We propose the novel task of automatic spectral calibration, aiming to advance the robustness of hyperspectral imaging in diverse real-world scenarios.

• Dataset Characteristics:

  • Camera: Specim IQ, featuring a spectral resolution of 3nm across the 400–1000nm range.
  • Recording Method: Each scene is captured twice:
    1. Without reference board: Captures raw scene data.
    2. With white reference board: Records illumination conditions under the same settings.
       This approach ensures asynchronous yet precise pairing of uncalibrated and calibrated HSIs, effectively minimizing illumination variability.
  • Dark Current Correction: Dark current noise, intrinsic to the camera sensor, is carefully recorded and subtracted during post-processing, ensuring high data accuracy.

• Scene Diversity:
  The dataset encompasses a wide range of urban and natural scenes, captured under various weather conditions, lighting scenarios, and times of day.

• Benchmarking Standard:
  BJTU-UVA establishes a new standard for spectral calibration by combining real-world scene variability with rigorous illumination recording, offering a robust foundation for testing and advancing spectral calibration techniques.

Dataset Splits

The file real_light_split.json provides predefined 'train', 'validation', and 'test' splits of the dataset, enabling standardized evaluation and comparison across different methods.

Synthetic Data Generation

The script gen_sim_data.py can be used to generate the BJTU-UVA-E dataset, a synthetic extension of BJTU-UVA, by simulating 10 different illumination conditions recorded in ill_sim.pkl. This allows researchers to explore and benchmark their methods under controlled, yet diverse, illumination variations.

Download

You can access the dataset via the following link:
BaiduNetDisk
Password: pgkg

Citation

@misc{du2024spectral,
title={Automatic Spectral Calibration of Hyperspectral Images: Method, Dataset and Benchmark},
author={Zhuoran Du and Shaodi You and Cheng Cheng and Shikui Wei},
year={2024},
eprint={2412.14925},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={ https://arxiv.org/abs/2412.14925 },
}

We hope BJTU-UVA inspires new research and development in the field of hyperspectral imaging. If you use this dataset, please consider citing our work.