Texts as Lines: Text Detection with Weak Supervision

Table 4

Experimental results for SCUT-TD500. “PT”refers to the model pretrained on SynthText. “Ext.” indicates external data. All compared methods were trained in a strongly supervised way. “FM” refers to the model trained with a full mask for strong-supervised learning.

Method	Ext.	MSRA-TD500
		Precision	Recall	F-measure	FPS
EAST [13]	—	87.3	67.4	76.1	13.2
RRPN [20]	—	82.0	68.0	74.0	—
DeepReg [18]	—	77.0	70.0	74.0	1.1
SegLink [40]		86.0	70.0	77.0	8.9
PixelLink [28]		83.0	73.2	77.8	3.0
RRD [26]		87.0	73.0	79.0	10
TextSnak [4]		83.2	73.9	78.3	1.1
[44]		87.6	76.2	82.9	—
TextField [29]		75.9	87.4	81.3	—
CRAFT [15]		88.2	78.2	82.9	8.6
MCN [45]		79.0	88.0	83.0	—
FM (ours)	PT	83.2	76.6	79.8	11.6
TAS (ours)	PT	80.6	74.1	77.2	12.0