A data set for color research

Kobus Barnard, Lindsay Martin, Brian Funt, Adam CoathColor 2002 · DOI

TL;DR

This paper introduces a comprehensive dataset for color research, particularly for computational color constancy and computer vision. It provides a rich collection of 743 images of scenes captured under 11 distinct illuminants, along with detailed camera characterization data and spectral information, enabling more robust and generalizable color research.

What Problem Does This Paper Address?

Accurate color perception and reproduction in digital imaging are challenging due to variations in illumination. Computational color constancy aims to correct for these variations, but progress has been hampered by a lack of standardized, comprehensive datasets that capture real-world complexity. Researchers often work with limited or proprietary data, making it difficult to compare methods or build upon previous work. This paper addresses this by releasing a large, well-documented dataset that includes diverse scenes, illuminants, and detailed camera calibration information, facilitating advancements in color constancy and related fields.

Methods

The authors meticulously collected and processed a large dataset. Key methodological aspects include:

  • Camera Characterization: They thoroughly characterized their Sony DXC-930 3CCD camera by measuring radiance spectra and corresponding camera responses using a Macbeth ColorChecker under various illuminant/filter combinations. This involved careful alignment of the camera and spectroradiometer, uniform illumination, and averaging to reduce noise.
  • Illuminant Spectra: Four sets of illuminant spectra were provided, ranging from common natural and artificial sources to more complex, measured ambient conditions. These were carefully selected to span a wide range of chromaticities.
  • Image Acquisition: Scenes were captured under 11 different illuminants. The process involved placing a reference white standard to estimate illuminant RGB, then capturing the scene itself. Extensive averaging (50 frames for camera, 20 cycles for spectroradiometer) was used to improve data quality and reduce noise.
  • Image Pre-processing: Images underwent several crucial steps: fixed pattern noise removal, linearization to correct for camera non-linearity and offset, and correction for spatially varying chromaticity shifts due to camera optics. This resulted in images where pixel intensity is proportional to scene radiance.
  • Data Organization: The dataset is organized into five main components: camera characterization data, spectral data (camera sensor responses, illuminant spectra, reflectance spectra), images of scenes, images of single objects, and fluorescent surface characterization data.

Datasets

The paper introduces a new, extensive dataset for color research. The core components include:

  • Scene Images: 743 images of scenes captured under 11 different illuminants, categorized into four groups based on specularities (minimal, dielectric, metallic) and the presence of fluorescent surfaces.
  • Object Images: Images of single objects on a black background, rotated under each illuminant, designed for image indexing experiments.
  • Spectral Data:
    • Camera sensor responses (red, green, blue) for the Sony DXC-930.
  • 4 sets of illuminant spectra.
  • Reflectance spectra from various sources, including Macbeth ColorChecker patches, Munsell chips, Dupont paint chips, natural objects, and surfaces from the Krinov dataset.
  • Camera Calibration Data: Data used to characterize the Sony DXC-930 camera, including radiance spectra and camera responses.
  • Fluorescent Surface Data: Input/output energy spectrum pairs for fluorescent surfaces, characterizing their unique behavior under different illuminants.

Key Insights and Ideas

  • The Importance of Contextualized Data: The authors strongly emphasize that data is only useful in context. Providing detailed information about the data collection process, including camera characterization and illuminant properties, is crucial for its effective use.
  • Bridging the Gap Between Theory and Practice: The dataset aims to bridge the gap between theoretical color constancy algorithms and their practical implementation by offering a rich set of real-world data that mimics complex lighting conditions.
  • Enabling Reproducibility and Collaboration: By making the data publicly available, the authors foster reproducibility of research and encourage collaborative progress in computational color science.
  • Handling Complex Surfaces: The inclusion of data for fluorescent surfaces and various types of specularities addresses challenging aspects of color perception that are often simplified or ignored in synthetic datasets.
  • High Dynamic Range Imaging: The pre-processing techniques allow for the study of higher dynamic range images, which are becoming more prevalent and offer richer information for analysis.

Key Results

The primary contribution is the creation and release of a comprehensive dataset. While the paper doesn't present quantitative performance metrics for specific algorithms, it highlights:

  • Dataset Size: 743 images of scenes, a significant number for color research.
  • Illuminant Diversity: Scenes captured under 11 carefully selected illuminants, covering a broad range of chromaticities.
  • Data Richness: The dataset includes not only images but also detailed spectral information and camera characterization data, providing a complete picture for color analysis.
  • Data Quality: Rigorous pre-processing steps (noise reduction, linearization, chromaticity correction) ensure high-quality, usable data.
  • Accessibility: The data is made available online, facilitating its widespread use by the research community.

Limitations and Future Work

  • Registration: The authors acknowledge that scene shift between illuminants might occur, meaning images are not absolutely guaranteed to be registered. However, they state that registration is "very close" and can be checked if required.
  • Illuminant Magnitude Error: The method for estimating illuminant RGB magnitude from white reflectance standards can have an error of "easily 10%" due to difficulties in maintaining precise perpendicularity to the light source.
  • Future Use: The authors anticipate a wide range of benefits for research into computational color science and computer vision, including color constancy algorithms, synthetic data experiments, and object recognition applications. They look forward to novel uses of the data.

Summary updated

Abstract

Abstract We present an extensive data set for color research that has been made available online ( www.cs. sfu.ca/∼colour/data) . The data are especially germane to research into computational color constancy, but we have also aimed to make the data as general as possible, and we anticipate a wide range of benefits to research into computational color science and computer vision. Because data are useful only in context, we provide the details of the collection process, including the camera characterization, and the data used to determine that characterization. The most significant part of the data is 743 images of scenes taken under a carefully chosen set of 11 illuminants. The data set also has several standardized sets of spectra for synthetic data experiments, including some data for fluorescent surfaces. © 2002 Wiley Periodicals, Inc. Col Res Appl, 27, 147–151, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/col.10049
BibTeX
@inproceedings{barnard2002data,
  author = {Kobus Barnard and Lindsay Martin and Brian Funt and Adam Coath},
  title = {{A data set for color research}},
  booktitle = {Color},
  year = {2002},
  doi = {10.1002/col.10049},
}
Figure