PST

Part-based Semantic Transform for Few-shot Semantic Segmentation

Overview

• networks/ contains the implementation of the PST(PST_net.py);
• models/ contains the backbone, semantic decomposition module;
• utils/ contains the semantic match module;

Dependencies

python == 3.7, pytorch1.0,

torchvision, pillow, opencv-python, pandas, matplotlib, scikit-image

Performance

Performance of k shot semantic segmentation on Pascal-5ⁱ

Setting	Backbone	Method	Pascal-50	Pascal-51	Pascal-52	Pascal-53	Mean
1-shot	VGG16	PST	48.89	64.65	51.44	47.52	53.12
	Resnet50	PMMs	51.98	67.54	51.54	49.81	55.22
		PST	52.66	67.13	53.23	51.48	56.14
5-shot	VGG16	PST	51.14	65.27	52.83	48.51	54.44
	Resnet50	PMMs	55.03	68.22	52.89	51.11	56.81
		PST	54.93	68.69	53.77	51.76	57.29

Performance of k shot semantic segmentation on MS COCO

Setting	Method	COCO-200	COCO-201	COCO-202	COCO-203	Mean
1-shot	PMMs	29.28	34.81	27.08	27.27	29.61
	PST	30.37	37.51	30.22	32.56	32.67
5-shot	PMMs	33.00	40.55	30.29	33.27	34.28
	PST	34.06	42.20	34.74	38.95	37.49