MIT Press

Table 3:

Content extraction performance for H-VILA. The H-VILA models significantly reduce the inference time cost compared to LayoutLM, while achieving comparable accuracy on the three benchmark datasets.

	GROTOAP2		DocBank		S2-VL		Inference Time (ms)
	Macro F1 $⇈$	H(G) $⇊$	Macro F1 $⇈$	H(G) $⇊$	Macro F1 $⇈$	H(G) $⇊$	Inference Time (ms)
LayoutLM_BASE	92.34	0.78	91.06	2.64	82.69(6.04)	4.19(0.25)	52.56(0.25)
Simple Group Classifier	92.65	0.00	87.01	0.00	–^¹	–	82.57(0.30)

H-VILA(Text Line)	91.65	0.32	91.27	1.07	83.69(2.92)	1.70(0.68)	28.07(0.37)^²
H-VILA(Text Block)	92.37	0.00	87.78	0.00	82.09(5.89)	0.36(0.12)	16.37(0.15)

	GROTOAP2		DocBank		S2-VL		Inference Time (ms)
	Macro F1 $⇈$	H(G) $⇊$	Macro F1 $⇈$	H(G) $⇊$	Macro F1 $⇈$	H(G) $⇊$	Inference Time (ms)
LayoutLM_BASE	92.34	0.78	91.06	2.64	82.69(6.04)	4.19(0.25)	52.56(0.25)
Simple Group Classifier	92.65	0.00	87.01	0.00	–^¹	–	82.57(0.30)

H-VILA(Text Line)	91.65	0.32	91.27	1.07	83.69(2.92)	1.70(0.68)	28.07(0.37)^²
H-VILA(Text Block)	92.37	0.00	87.78	0.00	82.09(5.89)	0.36(0.12)	16.37(0.15)

The simple group classifier fails to converge for one run. We do not report the results for fair comparison.

When reporting efficiency in other parts of the paper, we use this result because of its optimal combination of accuracy and efficiency.

View Large

This Feature Is Available To Subscribers Only