Texts as Lines: Text Detection with Weak Supervision

Table 1

Experimental results for Total-Text. “PT”refers to the model pretrained with one epoch on SynthText. “Ext.” indicates external data. “FM” refers to the model trained with full mask on Total-Text. All listed results were obtained in a strongly supervised manner.

Method	Ext.	Total-Text
		Precision	Recall	F-measure	FPS
SegLink [40]	—	30.0	23.8	26.7	—
EAST [13]	—	50.0	36.2	42.0	—
Mask TextSpotter [12]	—	69.0	55.0	61.3	—
PSENet [16]	—	81.8	75.1	78.3	3.9
[41]		80.9	76.2	78.5	10.0
TextSnake [4]		82.7	74.5	78.4	—
PSENet [16]		84.0	78.0	80.9	3.9
SPCNet [14]		83.0	82.8	82.9	—
CRAFT [15]		87.6	79.9	83.6	8.6
FM (ours)	PT	83.1	81.6	82.4	10.4
TAS (ours)	PT	78.5	76.7	77.6	11.2