Downstream Semantic Segmentation Model for Low-Level Surface Crack Detection

Table 1

Comparison of state-of-the-art deep learning models in four high-level semantic segmentation datasets. Note that numeric data filled in this table refers to mIOU evaluation results collected from original papers.

Model	Publication	Type	mIOU
			PASCAL VOC 2012	Pascal Context	Cityscapes	ADE20K
			(test set)	(val. set)	(test set)	(val. set)
PSPNet [30]	CVPR 2017	CNN	85.4	—	80.2	—
DeepLabV3 [29]	arXiv 2017	CNN	85.7	—	81.3	—
EncNet [31]	CVPR 2018	CNN	85.9	—	—	44.7
DeepLabV3+ [32]	ECCV 2018	CNN	87.8	—	82.1	–
Swin Transformer [33]	ICCV 2021	Transformer	—	—	—	53.5
Segmenter [34]	ICCV 2021	Transformer	—	59.0	—	51.8
SETR [35]	CVPR 2021	Transformer	—	55.8	81.6	50.3
SegFormer [36]	NeurIPS 2021	Transformer	—	—	83.1	51.8