Big Transfer Learning for Fine Art Classification

Table 3

State-of-the-art results for artist, genre, and style categorization on the WikiArt dataset, including samples and classes in each task.

Reference	Year	Method	Artist			Style			Genre
			Sample	Classes	Acc. (%)	Samples	Classes	Acc. (%)	Samples	Classes	Acc. (%)
[9]	2016	CNN fine-tuning (AlexNet)	19,050	23	76.11	81,444	27	54.5	64,993	10	74.14
[31]	2016	Feature fusion	18,599	23	63.06	78,449	27	45.97	63,691	10	60.28
[43]	2017	CNN fine-tuning (ResNet18)	17,100	57	77.7
[35]	2017	CNN fine-tuning (ResNet-34)							79,434	26	61.15
[44]	2018	CNN fine-tuning (CaffeNet)	20,320	23	81.94	96,014	27	56.43	86,087	10	77.6
[19]	2019	Two-stage classification approach				26,400	22	66.71
[32]	2020	RGB and brush stroke channels	9766	19	88.38	30,825	25	58.99	28,760	10	76.27
[34]	2021	Structure selection	19,050	23	91.73	81,444	27	69.97	64,993	10	78.03
BiT-S (ours)		Big transfer	19,050	23	91.34	81,444	27	68.27	64,993	10	80.88
BiT-M (ours)		Big transfer	19,050	23	93.50	81,444	27	71.24	64,993	10	82.39

The bold values show that the proposed transfer learning approach outperforms the previous work by a large margin and achieves state-of-the-art performance in the art field.