The State of the Art in Illumination Estimation:
A Survey and Evaluation

Abstract

Estimating the color of an illuminant solely from a single photograph has been a long standing challenge in many disciplines including computer graphics, computational photography, and computer vision. Although this problem is studied under different names in different contexts such as computational color constancy, illuminant estimation, and white balancing; they often refer to the same underlying phenomenon. This problem becomes inevitably more complicated when there is not one but multiple illuminants in a given scene. In the limit, each pixel can be assumed to have its own illuminant. Given the multi-faceted nature of this problem and many applications that it enables from simply producing better looking pictures to more accurate object detection, a multitude of illuminant estimation methods have evolved in the last decades. These methods range from traditional low-level heuristics, such as the gray world assumption, to biologically inspired models and learning-based approaches that have proliferated rapidly in recent years. Thus, the first purpose of this survey is to provide an in-depth look in this wide landscape of illumination estimation algorithms and provide an accurate taxonomy for them. Secondly, we propose a novel synthetic dataset generated from hyperspectral path tracing that simulates the output of more than a dozen cameras under varying illumination conditions. By using this dataset as well as several real-world datasets that are introduced for this purpose, we evaluate the performance of each algorithm using relevant metrics. To our knowledge, our study offers the most comprehensive survey and evaluation for illumination estimation algorithms.

CCS Concepts

• Computing methodologies → Computational photography
• Computing methodologies → Perception
• Computing methodologies → Ray tracing

Dataset

91 camera views × 15 cameras × 8 exposures = 10920 captures, each with its illuminant map.

Example Images

Download

The dataset (~50 GB) is split into 13 zip parts. Download all parts and extract them into the same folder.

Evaluation

Evaluation results of the algorithms on each dataset.