Big Transfer Learning for Fine Art Classification

Table 5

Comparison of the classification accuracies achieved with different initializations.

Weight initialization	Tasks/acc. (%)			Average
	Artist	Style	Genre
BiT-M	94.01	71.49	82.07	82.52
BiT-S	91.45	67.88	80.57	79.97
BiT-M-C	93.88	71.59	82.13	82.53
BiT-M-S	93.94	71.31	81.80	82.35
BiT-M-Mul	95.20	75.78	84.52	85.17

It shows BiT-M-Mul model achieves state-of-the-art performance compared with other weight initialization models.