This report independently applies the Quality Measures (QM) from ISO/IEC 5259-2:2024 to the WikiArt image dataset. We mapped DataClinic's 3-level diagnostic results to ISO QM criteria and independently interpreted and assigned verdicts for each item. Notably, this report critically identifies four discrepancies between DataClinic API text descriptions and actual chart data.

Com-ML-1 — Class Completeness: Caution

WikiArt covers 27 art movements, providing broad coverage of major Western art-historical periods. However, several rare movements have sample counts too low for ML training. Action Painting (98 images), Analytical Cubism (98), and Synthetic Cubism (120) fall far below the typical minimum training requirement of 300+ images for deep learning models. While all 27 movements are technically "present," some are effectively unlearnable, preventing a full Pass.

Con-ML-2 — Pixel Channel Consistency: Caution

The L1 pixel histogram (below) reveals dramatically different distributions across RGB channels. The Blue channel shows a pronounced left-skewed peak at 30-40, the Red channel exhibits a bimodal structure with a spike near 255, and the Green channel follows a relatively smooth mid-range distribution.

These distributions are art-historically explainable. The low Blue values originate from the brown grounds of traditional oil painting, while the Red 255 spike reflects cadmium red and vermillion pigments saturating in digital captures. Though artistically meaningful, this pattern signals the need for channel-specific normalization strategies in ML pipelines.

▲ L1 Pixel Histogram — Blue (left-skewed 30-40), Red (bimodal + 255 spike), Green (smooth mid-range). Dramatic cross-channel differences

Bal-ML-1 — 133x Class Imbalance

ISO 5259-2's Bal-ML-1 measures the degree of balance in per-class sample counts. A max-to-min ratio exceeding 10:1 is generally considered severe imbalance. WikiArt's Impressionism (13,060 images) versus Analytical Cubism (98 images) yields a ratio of 133:1 — roughly 7x more severe than the SpectralWaste recycling dataset's 19.6:1.

The more structural problem is that the standard deviation (3,269) exceeds the mean (3,016). This means the concept of an "average class" is essentially meaningless. The dataset is bifurcated into a few large movements (Impressionism, Realism, Romanticism) and many small ones (Cubism variants, Minimalism).

This imbalance is art-historically inevitable. Impressionism was a mass movement spanning decades across Europe with thousands of participating artists, while Analytical Cubism was a short-lived experiment (1907-1912) led by just two artists — Picasso and Braque. Yet art-historical inevitability does not excuse ML training problems. A model trained on this data will overfit to Impressionism while barely learning Analytical Cubism at all.

Eft-ML-1 — Classes Inseparable in L2

ISO 5259-2's Eft-ML assesses whether each class in the dataset can be distinguished through learning. In the L2 general-purpose lens (Wolfram ImageIdentify Net V2, 1,280-dim), the PCA plot and contour map show all 27 classes forming a single connected cloud. No inter-class boundaries are visible, meaning the general-purpose shape recognition AI cannot visually distinguish between art movements.

DataClinic rated L2 Geometry as "Good," which contradicts the chart data (see Discrepancy D3 below).

Acc-ML-7 — Label Accuracy: Fail

Two types of labeling errors were observed.

1. Systematic movement misclassification: Salvador Dali's works are labeled as Abstract Expressionism. Dali is unambiguously classified as Surrealism in art history, distinct from Abstract Expressionism in period, geography, and technique. This suggests that such errors may be systematic rather than isolated.

2. Medium contamination: The Pop Art class contains not only traditional paintings but also photographs of installations, architectural photography, and other non-painting media. The implicit assumption that "art = painting" breaks down in the Pop Art genre, which cascades into the extreme separation observed in L3 analysis.

Sim-ML-2 — Minimalism and Color Field Painting Merge

Sim-ML-2 measures cases where samples from different classes cluster too closely in embedding space. In the L2 analysis, Minimalism and Color Field Painting occupy nearly identical positions. These two movements are closely related in art history as well — both emerged in 1960s New York and share a focus on color planes and geometric simplicity — so the general-purpose AI's inability to separate them is somewhat expected.

From an ML standpoint, however, keeping these as separate classes means the classifier will fail to learn the boundary. Class merging or hierarchical labeling (e.g., Minimalism → "Geometric Abstraction" parent category) should be considered.

▲ L2 Contour — Two density centers within a single continuous mass. All 27 classes remain unseparated in one cloud

Rep-ML-1 — L2 Feature Space: Minimalism/Color Field Bias

The core of the L2 general-purpose lens's feature space (high-density region) is dominated by Minimalism and Color Field Painting. These two movements feature visually simple compositions (monochrome canvases, geometric forms) that the general-purpose shape recognition AI interprets as "the most universal visual patterns."

As a result, the rich visual diversity of 27 movements — Baroque's dramatic chiaroscuro, Ukiyo-e's woodblock textures, Expressionism's distorted forms — is inadequately represented in the feature space. This reflects both the limitations of the general-purpose lens and the dataset's representativeness deficit.

▲ L2 Box Chart — Per-class density distributions. Minimalism/Color_Field_Painting high-density concentration confirmed

Div-ML-1 — Diversity Failure in L2

ISO 5259-2's Div-ML-1 measures the effective dimensionality and distributional diversity of features. The fact that all 27 art movements collapse into a single continuous cloud under the L2 general-purpose lens means that, from this lens's perspective, "art movement diversity" does not exist in the data.

Wolfram ImageIdentify Net was trained for everyday object classification. To this model, all paintings are essentially one category: "image." The differences between movements — brushwork, palette, composition — register only as minuscule variations in the 1,280-dimensional space, insufficient for class separation.

▲ L3 Contour — Unlike L2, the BLIP lens reveals structure: dramatic Pop Art separation + period-based stratification