Edge Guided GANs with Contrastive Learning for Semantic Image Synthesis

Authors: Hao Tang, XIAOJUAN QI, Guolei Sun, Dan Xu, Nicu Sebe, Radu Timofte, Luc Van Gool

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on three challenging datasets show that our ECGAN achieves significantly better results than state-of-the-art methods.
Researcher Affiliation Academia 1ETH Zurich, 2University of Hong Kong, 3HKUST, 4University of Trento, 5University of Wurzburg
Pseudocode No The paper describes its methods and components in detail, often with block diagrams, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about providing open-source code for the proposed ECGAN method, nor does it provide a link to a code repository.
Open Datasets Yes We follow Gau GAN (Park et al., 2019) and conduct experiments on three datasets, i.e., Cityscapes (Cordts et al., 2016), ADE20K (Zhou et al., 2017), and COCO-Stuff (Caesar et al., 2018).
Dataset Splits No The paper mentions using training, validation, and test data in general terms but does not provide specific percentages or counts for how the datasets (Cityscapes, ADE20K, COCO-Stuff) are split for training, validation, and testing, nor does it explicitly reference predefined splits with specific details from cited works.
Hardware Specification Yes We conduct the experiments on an NVIDIA DGX1 with 8 V100 GPUs.
Software Dependencies No The paper mentions using components like Adam solver, Spectral normalization, Canny edge detector, and SegFormer, but does not provide specific version numbers for these or other software dependencies like programming languages or deep learning frameworks (e.g., PyTorch, TensorFlow, Python).
Experiment Setup Yes For both the image generator Gi and edge generator Ge, the kernel size and padding size of convolution layers are all 3x3 and 1 for preserving the feature map size. We set n=3 for generators Gi, Gs, and Gt. The channel size of feature F is set to C=64. ... We use the Adam solver (Kingma & Ba, 2015) and set β1=0, β2=0.999. λc, λs, λl, λf, and λp in Eq. 7 is set to 1, 1, 1, 10, and 10, respectively. All λ in both Eq. 6 and 7 are set to 2.