Weakly supervised video instance segmentation with scale adaptive generation regulation

模型	方法	AP/%	AP50/%	AP75/%	AR1	AR10	FPS
Baseline	Box-CrossVIS	33.0	61.5	32.3	35.4	39.0	18
Constant $α_{1}, α_{2}$	Box-CrossVIS+ARM（WBM： $α_{1} = 0.25, α_{2} = 1$ ）	34.4	66.7	32.6	34.9	40.4	18
	Box-CrossVIS+ARM（WBM： $α_{1} = 1, α_{2} = 0.25$ ）	34.8	64.3	35.4	34.9	39.9	17
	Box-CrossVIS+ARM（WBM： $α_{1} = α_{2} = 0.25$ ）	34.6	64.5	36.7	35.1	40.0	18
	Box-CrossVIS+ARM（WBM： $α_{1} = 0.50, α_{2} = 0.75$ ）	34.5	61.8	37.6	34.9	39.8	18
	Box-CrossVIS+ARM（WBM： $α_{1} = 0.75, α_{2} = 0.50$ ）	33.7	0.1	31.7	35.6	39.8	18
	Box-CrossVIS+ARM（WBM： $α_{1} = α_{2} = 0.50$ ）	34.1	62.2	34.8	35.7	39.9	17
	Box-CrossVIS+ARM（WBM： $α_{1} = α_{2} = 0.75$ ）	32.7	59.7	33.4	34.4	38.6	18
	Box-CrossVIS+ARM（WBM： $α_{1} = α_{2} = 1$ ）	34.1	64.8	34.9	35.3	40.5	18
Adaptive $α_{1}, α_{2}$	Box-CrossVIS+ARM（WUM）	33.4	60.2	33.6	35.4	39.6	17
Adaptive $α_{1}, α_{2}$	Box-CrossVIS+ARM（WNM）	35.6	64.3	36.8	36.7	41.1	18

Table 2. Comparison of effects of multi-scale fusion strategies on segmentation results

View table

View in Article

Table 2. Comparison of effects of multi-scale fusion strategies on segmentation results

层数	输入层	AP/%	AP50/%	AP75/%	AR1	AR10	FPS
基准	Box-CrossVIS	33.0	61.5	32.3	35.4	39.0	18
2	F3+F4	34.9	64.8	33.6	36.7	39.5	18
2	F6+F7	17.6	45.0	10.3	20.4	23.8	17
3	F3+F4+F5	35.6	64.3	36.3	36.5	40.5	18
4	F3+F4+F5+F6	34.6	62.2	32.5	36.2	39.8	17
5	F3+F4+F5+F6+F7	34.2	62.9	34.2	35.3	40.0	17

Table 3. Comparison of experimental results of different models

View table

View in Article

Table 3. Comparison of experimental results of different models

网络	弱监督	AP/%	AP50/%	AP75/%	AR1	AR10	FPS
Yolactedge^［20］		36.0	60.6	37.28	—	—	29
STMask^［18］		40.0	62.7	42.9	0.421	0.455	22
CrossVIS^［23］		40.6	65.7	42.2	43.2	46.4	16
Box-CrossVIS	√	33.0	61.5	32.3	35.4	39.0	18
SAGRNet	√	38.1	69.2	38.3	37.8	42.7	17

Table 4. Result of ablation experiments
View table
View in Article
Table 4. Result of ablation experiments
方法 ARM GRM AP/% AP50/% AP75/% AR1 AR10 FPS
Box-CrossVIS × × 33.0 61.5 32.3 35.4 39.0 18
Ours √ × 35.6 64.3 36.8 36.7 41.1 18
Ours × √ 35.6 64.3 36.3 36.5 40.5 18
Ours √ √ 38.1 69.2 38.3 37.8 42.7 17

Tools

Get Citation

Copy Citation Text

Yinhui ZHANG, Weiqi HAI, Zifen HE, Ying HUANG, Dongdong CHEN. Weakly supervised video instance segmentation with scale adaptive generation regulation[J]. Optics and Precision Engineering, 2023, 31(18): 2736

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites