Table 1 The architecture of G1/G2 network. “IN” represents InstanceNorm, “LReLU” represents Leaky ReLU activation, “Conv.”/“DeConv.” denotes convolutional/transposed convolutional layer with kernel size of 4, “st” means stride, “Concat” explains the skip connections, “Guidance” means guidance loss operation, and “Shift” means shift-connection operation. The different layers of G1 and G2 are listed separately
From: Stacked generative adversarial networks for image compositing
The generative model G1/G2 |
|---|
Input G1: Image (512×512×3)/ G2: Feature (512×512×6) |
Layer 1 G1: Conv. (3, 64), st=2;/ G2: Conv. (6, 64), st=2; |
Layer 2 LReLU; Conv.(64, 128), st=2; IN; |
Layer 3 LReLU; Conv.(128, 256), st=2; IN; |
Layer 4 LReLU; Conv.(256, 512), st=2;IN; |
Layer 5 LReLU; Conv.(512, 512), st=2; IN; |
Layer 6 LReLU; Conv.(512, 512), st=2; IN; |
Layer 7 LReLU; Conv.(512, 512), st=2; IN; |
Layer 8 LReLU; Conv.(512, 512), st=2; |
Layer 9 ReLU; DeConv.(512, 512), st=2; IN; Concat.(9, 7); |
Layer 10 ReLU; DeConv.(1024, 512), st=2; IN; Concat.(10, 6); |
Layer 11 ReLU; DeConv.(1024, 512), st=2; IN; Concat.(11, 5); |
Layer 12 ReLU; DeConv.(1024, 512), st=2; IN; Concat.(12, 4); |
Layer 13 ReLU; DeConv.(1024, 256), st=2; IN; Concat.(13, 3); |
Layer 14 ReLU; Guidance; Shift; DeConv.(768, 128), st=2; IN; Concat.(14, 2); |
Layer 15 ReLU; DeConv.(256, 64), st=2; IN; Concat.(15, 1); |
Layer 16 ReLU; DeConv.(128, 3), st=2; Tanh; |
Output G1: Feature (512×512×3)/ G2: Image (512×512×3) |