Semantic segmentation of textured mosaics

EURASIP Journal on Image and Video Processing

Table 6 Comparison of LBP + traditional classifier, SegNeXt, and DeepLabv3+ performance on 10-class test sets of SemSegOutex and SegSemDTD (standard and hard)

Metric	SemSegOutex 10-class			SemSegDTD 10-class			SemSegDTD 10-class (hard)
	LBP\(^{1}\)	SegNeXt	DLV3+	LBP\(^{2}\)	SegNeXt	DLV3+	LBP\(^{3}\)	SegNeXt	DLV3+
PA	0.9416	0.9959	0.9940	0.5715	0.9881	0.9924	0.2929	0.8706	0.8228
mPA	0.8635	0.9958	0.9939	0.5359	0.9884	0.9923	0.2713	0.8732	0.8244
mIoU	0.8122	0.9917	0.9880	0.3669	0.9775	0.9849	0.1650	0.7859	0.7088
\(\text {IoU}_1\)	0.8863	0.9915	0.9867	0.7088	0.9912	0.9896	0.0218	0.9071	0.6640
\(\text {IoU}_2\)	0.8951	0.9918	0.9894	0.4681	0.9233	0.9872	0.1512	0.9778	0.8735
\(\text {IoU}_3\)	0.8859	0.9922	0.9906	0.5510	0.9913	0.9908	0.0162	0.7952	0.5831
\(\text {IoU}_4\)	0.8688	0.9918	0.9871	0.2715	0.9921	0.9884	0.1221	0.5266	0.7100
\(\text {IoU}_5\)	0.8064	0.9916	0.9846	0.3230	0.9925	0.9832	0.1209	0.8389	0.7064
\(\text {IoU}_6\)	0.9412	0.9923	0.9905	0.4730	0.9924	0.9872	0.1052	0.4482	0.7827
\(\text {IoU}_7\)	0.8879	0.9914	0.9851	0.3933	0.9178	0.9886	0.4165	0.8677	0.6275
\(\text {IoU}_8\)	0.8868	0.9914	0.9847	0.1028	0.9921	0.9761	0.2857	0.5691	0.4594
\(\text {IoU}_9\)	0.9273	0.9917	0.9875	0.1674	0.9921	0.9852	0.3407	0.9890	0.9350
\(\text {IoU}_{10}\)	0.9005	0.9915	0.9936	0.5297	0.9907	0.9732	0.1875	0.9396	0.7462

Best metric values shown in bold. Figures for DLV3+ are repeated from Tables 1 and 2
\(^{1}\) Best LPB descriptor: \((P,R) = (24,3)\), \(HS = 24\); best classifier: medium Gaussian SVM
\(^{2}\) Best LPB descriptor: \((P,R) = (24,3)\), \(HS = 36\); best classifier: cubic SVM
\(^{3}\) Best LPB descriptor: \((P,R) = (24,3)\), \(HS = 48\); best classifier: weighted KNN