Answers are grounded in this paper and cite their sources.
Random splits of fine-art datasets that contain near-duplicate images inflate accuracy estimates, and ensemble stacking protocols rarely document whether meta-classifier training is free from leakage.We propose and apply a fully documented, leakage-controlled evaluation protocol to the Best Artworks of All Time dataset.Nearduplicates are removed with perceptual hashing (pHash, Hamming ≤ 10; 547 groups, 595 images removed) before any split, and the deduplication is audited on both sides of the threshold (338 pairs; precision 0.97, recall 0.95), which bounds residual train-test duplication at 0.4-1.2% of test images per seed.A seed-matched comparison with nearduplicates retained raises Xception accuracy by 2.47 percentage points (95% CI [0.28, 4.66]), reported as a protocolassociated change rather than as an isolated estimate of cross-split leakage.With the test set held fixed, near-duplicates added only to the training split do not measurably raise accuracy (-0.14 percentage points, 95% CI [-1.58, +1.30], five seeds), and a duplicate-group-aware split that keeps all 4,388 images gives the same accuracy as the deletion policy (+0.42 percentage points, 95% CI [-0.42, +1.26]).As a reference model, four CNN backbones are combined by five-fold out-of-fold stacking with early stopping on an inner validation subset, never on the held-out fold.Over five seeds the ensemble reaches 89.63% (±1.85%) accuracy; by McNemar's test it exceeds ResNet50 in 4 of 5 seeds (3 after Holm correction) and Xception in 3 of 5 (2 after Holm), but not equal-weight soft voting in any seed, and a two-backbone stack reproduces it (89.84%± 1.76%).Under the same protocol Swin-Tiny exceeds the ensemble in every seed (93.74% ± 0.66%; significant in 5 of 5 seeds, 3 after Holm) and ConvNeXtTiny is statistically indistinguishable from it (90.05%± 0.51%); these baselines are architectural representatives of recent methods, not reproductions of them.The contribution is the protocol and the resulting multi-seed benchmark rather than any single model.The conclusions are limited to the defined ten-class benchmark of this collection.
No takes yet. Share an insight, caveat, or question.
Haitham Qutaiba Ghadhban (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: