MIT Press

Table 5:

Comparison between I-VILA models and other layout-aware methods that require expensive pretraining. I-VILA achieves comparable accuracy with less than 5% of the training cost.

	GROTOAP2		DocBank		S2-VL		Training Cost^¹
	F1 $⇈$	H(G) $⇊$	F1 $⇈$	H(G) $⇊$	F1 $⇈$	H(G) $⇊$	Training Cost^¹
BERT_BASE (Devlin et al., 2019)	90.78	1.58	87.24	3.50	78.34(6.53)^¹	7.17(0.95)	40 hr fine-tuning
BERT_BASE + I-VILA(Text Line)	91.65	1.13	90.25	2.56	81.15(4.83)	4.76(1.28)	40 hr fine-tuning
BERT_BASE + I-VILA(Text Block)	92.31	0.63	89.49	2.25	81.82(4.88)	3.65(0.26)	40 hr fine-tuning

LayoutLM_BASE (Xu et al., 2020)	92.34	0.78	91.06	2.64	82.69(6.04)	4.19(0.25)	1.2k hr pretraining + 50 hr fine-tuning
LayoutLMv2_BASE (Xu et al., 2021)	–^²	–	93.33	1.93	83.05(4.51)	3.34(0.82)	9.6k hr pretraining^³ + 130 hr fine-tuning

	GROTOAP2		DocBank		S2-VL		Training Cost^¹
	F1 $⇈$	H(G) $⇊$	F1 $⇈$	H(G) $⇊$	F1 $⇈$	H(G) $⇊$	Training Cost^¹
BERT_BASE (Devlin et al., 2019)	90.78	1.58	87.24	3.50	78.34(6.53)^¹	7.17(0.95)	40 hr fine-tuning
BERT_BASE + I-VILA(Text Line)	91.65	1.13	90.25	2.56	81.15(4.83)	4.76(1.28)	40 hr fine-tuning
BERT_BASE + I-VILA(Text Block)	92.31	0.63	89.49	2.25	81.82(4.88)	3.65(0.26)	40 hr fine-tuning

LayoutLM_BASE (Xu et al., 2020)	92.34	0.78	91.06	2.64	82.69(6.04)	4.19(0.25)	1.2k hr pretraining + 50 hr fine-tuning
LayoutLMv2_BASE (Xu et al., 2021)	–^²	–	93.33	1.93	83.05(4.51)	3.34(0.82)	9.6k hr pretraining^³ + 130 hr fine-tuning

We report the equivalent V100 GPU hours on the GROTOAP dataset in this column.

LayoutLMv2 cannot be trained on the GROTOAP2 dataset because almost 30% of its instances do not have compatible PDF images.

The authors do not report the exact cost in the paper. The number is a rough estimate based on our experimental results.

View Large

This Feature Is Available To Subscribers Only