Dual conditional GAN based on external attention for semantic image synthesis

Figures & data

Figure 1. Architecture of DCSIS.

Figure 2. Architecture of the generator network.

Figure 3. Architecture of the RGB semantic encoder.

Figure 4. Graph attention module.

Figure 5. Graph convolution segmentation network.

Table 1. Quantitative comparison with competing methods on different datasets.

		Cityscapes		ADE20K		CelebAMask-HQ
Methods	#param	FID↓	mIoU↑	FID↓	mIoU↑	FID↓	mIoU↑
Pix2pixHD	183M	94.5	57.8	80.8	19.5	25.8	71.2
SPADE	102M	58.6	61.2	32.6	37.2	23.1	72.1
CRN	89M	102.1	51.7	71.8	21.4	29.8	72.2
CC-FPSE	132M	52.8	62.9	32.6	41.3	21.1	n/a
SCGAN	–	49.4	64.1	32.1	43.8	19.2	73.3
SDM	–	52.3	66.4	33.6	38.4	22.4	75.3
OASIS	96M	47.8	65.8	29.2	45.3	17.3	75.1
Ours	110M	48.4	68.1	31.8	47.4	17.1	77.8

Figure 6. Visual comparison of CelebAMask-HQ dataset.

Figure 7. Visual comparison of Cityscapes dataset.

Figure 8. Visual comparison of ADE20K dataset.

Figure 9. Visual comparison of different variants.

Table 2. Quantitative comparison of ablation experiments on Cityscapes and CelebAMask-HQ datasets.

	Cityscapes		CelebAMask-HQ
Methods	FID↓	mIoU↑	FID↓	mIoU↑
Baseline	59.1	61.2	23.5	70.8
+Unet	49.2	65.2	18.9	74.9
+GSeg	48.9	66.5	17.6	76.8
+DCA + GSeg	46.9	68.8	17.2	78.9
+DCA + GSeg + GAT	46.6	69.6	16.8	79.3