Table 7 The segmentation accuracy and... | Emerald Publishing

Table 7

The segmentation accuracy and efficiency of lightweight MOS methods on YouTube-VOS dataset.

					Seen		Unseen
No.	Year	Method	GPU	J&F Mean (%)↑	J Mean (%)↑	F Mean (%)↑	J Mean (%)↑	F Mean (%)↑	Inference Speed (fps) ↑	# Param (M) ↓	Model Size (MB) ↓
1	2020	FRTM-fast (ResNet-18) [170]	Tesla V100	65.7	68.6	71.3	58.4	64.5	41.3^*	13.2	53.9
2	2021	AOT-T [232]		79.7	79.6	83.8	73.7	81.8	41	5.7	23.3
3	2020	FRTM [170]		72.1	72.3	76.2	65.9	74.1	21.9^*	46.9	187.6
4	2021	RM Net [225]		81.5	82.1	85.7	75.7	82.4	12^*	53	202
5	2020	SAT [32]	RTX 2080 Ti	63.6	67.1	55.3	70.2	61.7	39^*	72	288
6	2021	TVOS [247]	Titan Xp	67.8	67.1	69.4	63	71.6	37^**	12.7	50
7	2019	AGSS [112]		71.3	71.3	75.2	65.5	73.1	12.5	48	197
8	2019	SiamMask(ResNet-50) [210]	RTX 2080	52.8	60.2	58.2	45.1	47.7	55^*	27	105.9
9	2021	SwiftNet(ResNet-18) [206]	Tesla P100	73.2	73.3	76.3	68.1	75	70^**	32.5	130
10	2021	SwiftNet(ResNet-50) [206]		77.8	77.8	81.8	72.3	79.5	25^**	37.4	149
11	2020	LSTNet [207]	Titan X Pascal	71.8	70.9	74.9	66.8	74.8	12.8	-	-
12	2020	GC [105]	Tesla P40	73.2	72.6	75.6	68.9	75.7	25^*	45.9	175
13	2021	G-FRTM-fast (RN18, τ =1) [155]	-	60.9	65.1	66.7	53	58.8	37.6^*	-	-
14	2020	MSN [222]	-	71.1	72.4	75.2	65.4	71.4	10^*	-	-

					Seen		Unseen
No.	Year	Method	GPU	J&F Mean (%)↑	J Mean (%)↑	F Mean (%)↑	J Mean (%)↑	F Mean (%)↑	Inference Speed (fps) ↑	# Param (M) ↓	Model Size (MB) ↓
1	2020	FRTM-fast (ResNet-18) [170]	Tesla V100	65.7	68.6	71.3	58.4	64.5	41.3^*	13.2	53.9
2	2021	AOT-T [232]		79.7	79.6	83.8	73.7	81.8	41	5.7	23.3
3	2020	FRTM [170]		72.1	72.3	76.2	65.9	74.1	21.9^*	46.9	187.6
4	2021	RM Net [225]		81.5	82.1	85.7	75.7	82.4	12^*	53	202
5	2020	SAT [32]	RTX 2080 Ti	63.6	67.1	55.3	70.2	61.7	39^*	72	288
6	2021	TVOS [247]	Titan Xp	67.8	67.1	69.4	63	71.6	37^**	12.7	50
7	2019	AGSS [112]		71.3	71.3	75.2	65.5	73.1	12.5	48	197
8	2019	SiamMask(ResNet-50) [210]	RTX 2080	52.8	60.2	58.2	45.1	47.7	55^*	27	105.9
9	2021	SwiftNet(ResNet-18) [206]	Tesla P100	73.2	73.3	76.3	68.1	75	70^**	32.5	130
10	2021	SwiftNet(ResNet-50) [206]		77.8	77.8	81.8	72.3	79.5	25^**	37.4	149
11	2020	LSTNet [207]	Titan X Pascal	71.8	70.9	74.9	66.8	74.8	12.8	-	-
12	2020	GC [105]	Tesla P40	73.2	72.6	75.6	68.9	75.7	25^*	45.9	175
13	2021	G-FRTM-fast (RN18, τ =1) [155]	-	60.9	65.1	66.7	53	58.8	37.6^*	-	-
14	2020	MSN [222]	-	71.1	72.4	75.2	65.4	71.4	10^*	-	-

Note:^* The inference speeds of these models were evaluated on the DAVIS 2016 dataset in the original papers.

**

The inference speeds of these models were evaluated on the DAVIS 2017 dataset in the original papers.