Color-in-Context: A 12K-Image Dataset for Color Recognition Under Varied Illumination

1. Introduction

Perceived color depends on surface reflectance, scene illumination, and sensor characteristics. Although large-scale benchmarks (e.g., ImageNet, COCO) have advanced general-purpose computer vision, illumination is rarely encoded as an explicit, controllable factor [1,2]. This can complicate the study of color recognition in settings where lighting conditions vary across capture environments.

Color constancy research seeks to offset the effects of scene illumination so that the perceived appearance of objects remains consistent across lighting conditions. Widely used benchmark resources—such as the reprocessed Gehler–Shi (ColorChecker) dataset, the NUS 8-Camera dataset, and Intel-TUT—provide calibrated images and standardized evaluation protocols for estimating the illuminant [3,4,5]. However, these datasets primarily target illuminant estimation and classical color-constancy objectives, rather than supervised recognition tasks defined over an explicit joint label space of color × illumination.

A related and persistent concern in computer vision is robustness under distribution shift. Benchmarks such as ImageNet-C and ImageNet-P quantify performance degradation under common corruptions and perturbations [6], while domain generalization literature surveys augmentation- and alignment-based approaches aimed at improving transfer to unseen conditions [7,8]. Because illumination changes are among the most prevalent sources of real-world shift, datasets that treat lighting variation as a first-class annotation are especially valuable for systematic evaluation.

Color-in-Context addresses this need by organizing images into 36 categories (9 colors × 4 illumination conditions). The dataset contributes: (i) a clear directory structure with standardized labels, (ii) dataset-level distribution statistics, (iii) recommended procedures for generating train/validation/test splits over the joint label space, and (iv) reference settings for baseline training and reporting.

2. Related Work

Classical color-constancy methods (gray-world, max-RGB, shades-of-gray) infer illuminant properties from image statistics [9,10,11], while probabilistic formulations incorporate priors over reflectance and lighting [12]. Deep learning approaches treat illuminant estimation as supervised regression or classification, typically trained on calibrated corpora such as ColorChecker and NUS [13,14]. In parallel, color naming studies highlight context dependence and category boundary effects that can influence perceived color under different illumination [15,16].

Robustness work documents performance degradation under corruptions; methods such as RandAugment, AugMix, and test-time adaptation have been explored to mitigate these effects [6,17,18,19]. Domain adaptation and generalization methods address transfer through alignment, reweighting, meta-learning, and distributionally robust objectives [7,8,20,21,22]. Practical evaluation often emphasizes macro-averaged metrics and explicit imbalance handling via reweighting, focal loss, or resampling [23,24,25,26]. A dataset that annotates both color and illumination offers a structured resource for studying these practices under lighting variation.

3. Dataset Design

3.1. Label Space

Colors (9): Black, Blue, Gray, Orange, Pink, Purple, Skyblue, White, Yellow.

Illuminations (4): fluorescentLight, indoor, indoorNight, sunLight.

Labels are encoded as <Color>/<Illumination>/filename (Figure 1).

3.2. Rationale

The illumination categories reflect common capture environments: typical indoor lighting, fluorescent lighting, indoor scenes at night, and outdoor scenes under sunlight. The color categories are frequent and practically useful; “Skyblue” is retained as a distinct label to match common usage in collected annotations.

3.3. Quality Controls

We applied vocabulary and case validation, perceptual-hash screening for near-duplicate detection, and optional EXIF capture when present. Images are treated as sRGB unless otherwise specified; any color-space conversions should be documented because they may affect chromaticity.

4. Data Summary and Distribution

The dataset exhibits non-uniform sampling across both annotation axes. Table 1 summarizes counts by illumination, Table 2 summarizes counts by color, and Table 3 lists selected joint buckets with notable counts. These summaries document dataset composition and support controlled sampling decisions (e.g., stratified splits) during downstream use. Figure 2 shows example images for each color category under each illumination condition.

5. Data Records

5.1. Access and Organization

The dataset is distributed via Kaggle as a versioned release (v2) at https://www.kaggle.com/datasets/nizamuddinmaitlo/different-colors-in-challenging-lightening-v2. Images are organized using the directory structure <Color>/<Illumination>/filename. Each file belongs to exactly one of the 36 joint categories defined by the cross-product of the nine color labels and four illumination labels.

5.2. Manifest and Metadata

The release may include an optional manifest file (e.g., manifest.csv) listing, for each image, its relative path and labels. Where available, additional fields may include file size and a cryptographic checksum (e.g., SHA-256) to support integrity verification. Optional fields such as image width and height may be included when generated during curation.

5.3. Versioning and Label Normalization

This manuscript describes the Kaggle dataset release labeled v2. A minor label inconsistency was normalized during curation (e.g., sunlight to sunLight); any future changes are tracked through Kaggle’s versioning mechanism and documented on the dataset page.

6. Technical Validation

We performed validation steps focused on label consistency, file integrity, and reduction of near-duplicate content.

6.1. Label Consistency Checks

Directory names and labels were normalized to a fixed vocabulary and case convention. Automated checks verified that each image path matches the expected <Color>/<Illumination>/ pattern and that labels belong to the predefined sets. Deviations (e.g., misspellings, mixed case) were corrected during normalization.

6.2. Duplicate and Near-Duplicate Screening

We applied perceptual-hash screening to identify exact and near-duplicate images within and across buckets. Candidate duplicates were reviewed and removed when appropriate, retaining a single representative instance per duplicated group.

6.3. File Integrity

Where available, checksums provided in the manifest can be used to verify integrity after download. We additionally checked that images are readable and decodable using standard image libraries; files failing decoding or exhibiting corruption were removed or replaced before release.

6.4. Manual Inspection

A manual inspection pass was conducted on a small, randomly sampled subset drawn across multiple buckets to confirm that directory labels correspond to the intended color and illumination tags and to identify obvious annotation errors. Identified inconsistencies were corrected during curation.

7. Usage Notes

7.1. Recommended Splits

For supervised learning experiments, we recommend generating splits stratified over the joint label (color × illumination) to preserve bucket proportions. A practical default is 70/15/15 train/validation/test, with minimum per-bucket counts enforced in validation and test. For illumination-transfer evaluation, users may construct splits that hold out one or more illumination conditions during training (e.g., train on indoor+fluorescentLight and evaluate on sunLight+indoorNight), ensuring that held-out conditions are not used for model selection.

7.2. Metrics and Reporting

We recommend macro-averaged metrics (macro-accuracy and macro-F1) across all 36 buckets, along with per-illumination and per-color breakdowns. Confusion matrices stratified by illumination can be reported. If probabilistic outputs are used, calibration measures such as expected calibration error (ECE) may be included.

7.3. Imbalance-Aware Training

Because bucket counts are not uniform (Table 3), users may consider imbalance-handling strategies such as inverse-frequency weighting at the bucket level, focal loss, or controlled resampling. Any resampling strategy should be documented to preserve reproducibility.

7.4. Data Augmentation Considerations

Geometric augmentation (e.g., flips and crops) is generally compatible with the label space. Photometric augmentation should be used conservatively because strong color jitter can directly alter the target label. If photometric augmentation is applied, users should document parameter ranges and verify that augmentations do not systematically change perceived color categories.

7.5. Reference Model Configurations

Users may evaluate standard architectures such as MobileNetV3, ResNet-18, and ViT-Tiny using ImageNet-pretrained initialization and $224\times 224$ inputs. Common regularization choices include early stopping based on a validation metric (e.g., macro-F1) and weight decay in the range $1\times {10}^{-4}$ to $5\times {10}^{-4}$. If class weights are used, bucket-level inverse-frequency weighting is a simple baseline. Augmentation should remain conservative with respect to color labels.

7.6. Reproducible Data Access

We recommend using the provided folder structure and (if present) the manifest file to load data deterministically. When creating derived subsets or splits, saving the corresponding file lists and random seeds supports reproducible experimentation.

8. Ethical Considerations

Perceived color can vary across sensors, devices, and display settings; users should document preprocessing and color-space assumptions. The dataset should be curated to avoid personally identifiable information. Downstream users are encouraged to assess potential contextual leakage (e.g., background correlations) and to document label definitions, noting that color naming may vary across contexts; “Skyblue” is retained as a practical descriptor aligned with collected annotations.

9. Limitations and Future Work

The dataset covers four illumination types and nine color categories, but it does not explicitly control for material properties (e.g., matte vs. glossy), cast shadows, or complex outdoor lighting. Future releases may expand coverage for sunLight, diversify materials and textures, and include additional annotations (e.g., object masks) where feasible.