Source-linked AI summary
On Finding Gray Pixels
Yanlin Qian, Joni-Kristian Kämäräinen, Jarno Nikkanen, Jiri Matas
TL;DR
Finding gray pixels in color-biased images is difficult because surface color and illumination can be ambiguous. This paper introduces a learning-free Grayness Index based on the Dichromatic Reflection Model to estimate global or spatial illumination. GI outperforms statistical methods and many deep methods on the reported benchmarks, with cross-dataset results outperforming all compared methods.
Problem
Finding achromatic pixels in color-biased images is ill-posed because different surface and illumination combinations can produce similar observed colors.
Source: §1 · p. 1Method
GI uses the Dichromatic Reflection Model to rank pixels by grayness, then applies clustering to estimate spatial illumination when multiple light sources are present.
Source: §1 · p. 2, §3.3 · p. 4Results
GI outperforms all compared learning-based and statistical methods in the cross-dataset setting and outperforms several popular learning-based methods in the single-dataset setting.
Source: §4.1 · p. 6, §4.2 · p. 7Takeaways & Limitations
GI estimates illumination on standard benchmarks more accurately than state-of-the-art learning-free methods in about 0.4 seconds.
Source: §6 · p. 8When the Macbeth Color Checker lies mainly under one illumination, its measurements can bias the ground truth and restrict GI's performance ceiling.
Source: §5 · p. 8
Abstract
from arXiv · showhide
We propose a novel grayness index for finding gray pixels and demonstrate its effectiveness and efficiency in illumination estimation. The grayness index, GI in short, is derived using the Dichromatic Reflection Model and is learning-free. GI allows to estimate one or multiple illumination sources in color-biased images. On standard single-illumination and multiple-illumination estimation benchmarks, GI outperforms state-of-the-art statistical methods and many recent deep methods. GI is simple and fast, written in a few dozen lines of code, processing a 1080p image in ~0.4 seconds with a non-optimized Matlab code.
1. Introduction
Color constancy seeks to adapt camera colors to illumination, but finding achromatic pixels in color-biased images is ill-posed. This paper proposes a learning-free Grayness Index (GI), derived from the Dichromatic Reflection Model, to rank pixels by grayness and estimate illumination.
1. Introduction
- Color constancy is ill-posed in color-biased images because an observed color can arise from different surface and illumination combinations.For example, cyan pixels could be a white surface under cyan light or a cyan surface under white light.Source: §1 · p. 1
- Learning-free methods estimate illumination per image, while learning-based methods depend on training and test images sharing illumination or scene statistics.The paper motivates learning-free approaches for correcting images from cameras unseen during training.Source: §1 · p. 1
- GI derives a physically interpretable grayness ranking from the Dichromatic Reflection Model and supports global or pixel-level illumination estimation.The introduction also reports simple, fast computation and consistent predictions across cameras.Source: §1 · pp. 1–2
- GI is reported to outperform state-of-the-art learning-free methods and achieve state-of-the-art cross-dataset results on single- and multi-illumination benchmarks.The paper presents GI as simple and fast, with a clear physical meaning and some qualitative ability to handle specular highlights.Source: §1 · p. 2
2. Related Work
Related work frames color constancy as estimating illumination from an image, using learned image-to-illumination mappings, statistical regularities, or physical image-formation models. GI draws on the Dichromatic Reflection Model but is classified as statistical because it finds gray pixels using image statistics.
2. Related Work
- The Dichromatic Reflection Model represents a pixel as body and surface reflections under global illumination, with the image-formation equation simplified under a narrow spectral-response assumption.The model's chromatic terms describe body and surface reflection, while its achromatic terms give their intensities.Source: §2 · p. 2
- Learning-based methods infer illumination from extensive training data, whereas learning-free methods use assumptions about local or global regularities in illumination and reflectance.The paper notes that learned methods depend heavily on overlap in cameras or scenes between training and test data.Source: §2 · p. 2
- Physics-based methods model image formation and can represent highlights and inter-reflections, but many rely on dichromatic-line intersections that often fail on natural images.GI uses the Dichromatic Model but is classified as statistical because its core is finding gray pixels from image statistics.Source: §2 · p. 2
- Xiong et al. search for gray surfaces in a camera-dependent LIS space, while Gray Pixel is identified as GI's closest related method.The paper says Gray Pixel is outlined in detail later.Source: §2 · p. 2
3. Grayness Index
The section analyzes why the earlier Gray Pixel constraints can mark colored pixels as gray, then introduces a DRM-based grayness index. It applies GI-ranked pixels to estimate either global or spatially varying illumination.
3. Grayness Index
- The earlier Gray Pixel method is revisited in the context of the Dichromatic Reflection Model.Source: §3 · p. 3
3.1. Gray Pixel in [41]
- The earlier Gray Pixel constraints can misidentify colored pixels: surface-reflection interaction can create false gray detections, while varying illumination on a homogeneous gray surface can yield equal channel contrast.Source: §3.1 · p. 3
3.2. Grayness Index using Dichromatic Reflection Model
- The DRM-based GI ranks pixels by grayness, with smaller values indicating pixels more likely to be gray.Unlike GP, GI uses a more complete image-formation model and does not weaken dark pixels according to luminance.Source: §3.2 · pp. 3–4
- GI computation applies a spatial-cue restriction and averages the map over a 7 × 7 window to weaken isolated detections attributed mainly to camera noise.Figure 2 traces the grayness map and selected pixels through illumination estimation and color correction.Source: §3.2 · pp. 3–4, Figure 2 · p. 3
3.3. GI Application in Color Constancy
- For one global light, the estimated illumination is the average color of the top N% of pixels ranked by GI.Source: §3.3 · p. 4
- For multiple lights, GI selects the top N% of pixels for K-means clustering, averages each cluster, and forms a spatial illumination map that favors nearby pixels sharing illumination.Source: §3.3 · p. 4
4. Evaluation
The evaluation tests GI in single- and multi-illumination settings, including cross-dataset transfer and parameter sensitivity. GI performs competitively across these settings, with reported strengths in unseen-camera evaluation and multi-illumination estimation.
4. Evaluation
- The evaluation covers single-illumination, multi-illumination, and cross-dataset settings, using angular error to assess estimated against ground-truth illumination.The datasets include Gehler-Shi, NUS 8-Camera, and MIMO.Source: §4 · p. 4
4.1. Single-dataset Setting
- In the single-dataset setting, GI beats several popular learning-based methods, although FFCC achieves the best overall performance on both datasets.The comparison includes methods from Gamut to Corrected-Moment; GI requires no extensive training or parameter tuning.Source: §4.1 · pp. 5–6
4.2. Cross-Dataset Setting
- GI outperforms all tested learning-based and statistical methods when training and test datasets have no cameras in common.The cross-dataset comparison uses each method’s best or final setting.Source: §4.2 · p. 7
- GI achieves more stable results across the NUS 8-Camera Dataset, with smaller standard deviation across cameras and lower sensitivity to camera hardware.The evaluation leverages the dataset’s shared scenes across eight cameras.Source: §4.2 · p. 7
4.3. Grid Search on Parameters
- N = 1e−1 and ϵ = 1e−4 give a good trade-off between mean and median angular error on Gehler-Shi and NUS 8-camera.The parameter grid search indicates robustness to tuning across orders of magnitude.Source: §4.3 · p. 7, Figure 5 · p. 7
4.4. Multi-illumination Setting
- On MIMO, GI outperforms methods despite not being designed for spatial illumination changes, with more clusters helping indoors but not in wild images.K-means predictions can be less sharp in scenes with complex geometry, while overall error and visual correction remain plausible.Source: Table 3 · p. 6, §4.4 · p. 7
5. Problems with the “Ground-truth”
GI’s angular error changes with the illumination setup and the size of the MCC-centered test region. In two-illumination images, an MCC dominated by one illuminant may bias the reference and limit GI’s measured performance.
5. Problems with the “Ground-truth”
- Some images contain gray pixels under two illuminants; GI can identify them, yet this can coincide with a large angular error.The authors hypothesize that the MCC is more dominated by one illuminant in these cases.Source: §5 · p. 8
- The experiment compares GI and FFCC on 66 two-illumination and 502 single-illumination Gehler-Shi images.The images are cropped to boxes centered at the MCC, with box sizes decreasing from A to E.Source: §5 · p. 8
- GI’s angular error increases as the test box shrinks for single-illumination images, but the trend reverses for double-illumination images.The authors suggest that an MCC placed mainly under one illuminant may create a biased ground truth that limits GI and possibly other statistical methods.Source: §5 · p. 8
6. Conclusions
The paper presents GI, a Dichromatic Reflection Model-based method for grayness estimation, and reports accurate, efficient illumination estimation on single- and multi-illumination tasks. It also concludes that learning-based methods generally perform worse in cross-dataset evaluation and can exploit color-checker location cues even when the checker is masked.
6. Conclusions
- GI estimates illumination more accurately than state-of-the-art learning-free methods on standard benchmarks, with processing time of about 0.4 seconds.The authors report effectiveness and efficiency on single- and multi-illumination estimation tasks.Source: §6 · p. 8
- Learning-based methods generally perform worse in cross-dataset evaluation and can exploit color-checker location cues when the checker is masked by zeros.The authors report that these methods can overfit to scene- and camera-specific features.Source: §6 · p. 8