![\"\"](\"https://media-exp1.licdn.com/dms/image/C5622AQF7Tj47ZkAORA/feedshare-shrink_800-alternative/0/1610864587761?e=1613606400&v=beta&t=yoEaZnYJ6lqNZxbemX1ID4cNMCv8nFFfqntBFYit2SM\")
You Only Need Adversarial Supervision for Semantic Image Synthesis • Despite their recent successes, GAN models for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision.
\nHistorically, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limiting the progress of GAN models for semantic image synthesis.
\nIn this work, they propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results.
\nThey re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training.
\nBy providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, they are able to synthesize images of higher fidelity with better alignment to their input label maps, making the use of the perceptual loss superfluous.
\nAuthors -> \nVadim Sushko, Edgar Schoenfeld, Dan Zhang, Juergen Gall, Bernt Schiele, Anna Khoreva
", "messages": [ { "id": "1156464", "postDate": "01/17/2021 07:07:27", "content": "You Only Need Adversarial Supervision for Semantic Image Synthesis • Despite their recent successes, GAN models for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision.
\nHistorically, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limiting the progress of GAN models for semantic image synthesis.
\nIn this work, they propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results.
\nThey re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training.
\nBy providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, they are able to synthesize images of higher fidelity with better alignment to their input label maps, making the use of the perceptual loss superfluous.
\nAuthors -> \nVadim Sushko, Edgar Schoenfeld, Dan Zhang, Juergen Gall, Bernt Schiele, Anna Khoreva
", "rawMarkdown": "You Only Need Adversarial Supervision for Semantic Image Synthesis • Despite their recent successes, GAN models for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision.\n\n\n\n# GitHub https://github.com/boschresearch/OASIS\n# Paper https://arxiv.org/abs/2012.04781\n\nHistorically, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limiting the progress of GAN models for semantic image synthesis.\n\nIn this work, they propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results.\n\nThey re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training.\n\nBy providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, they are able to synthesize images of higher fidelity with better alignment to their input label maps, making the use of the perceptual loss superfluous.\n\n**Authors -> \nVadim Sushko, Edgar Schoenfeld, Dan Zhang, Juergen Gall, Bernt Schiele, Anna Khoreva**", "votes": null }, { "id": "1157002", "postDate": "01/17/2021 15:24:14", "content": "Thanks for sharing @mobassir
", "rawMarkdown": "Thanks for sharing @mobassir", "votes": null } ], "comments": [ { "id": 1157002, "author_name": "sadmanaraf", "author_url": "", "post_date": "01/17/2021 15:24:14", "content": "Thanks for sharing @mobassir
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1156464": "You Only Need Adversarial Supervision for Semantic Image Synthesis • Despite their recent successes, GAN models for semantic image synthesis still suffer from poor image quality when trained with only adversarial supervision.\n\n\n\n# GitHub https://github.com/boschresearch/OASIS\n# Paper https://arxiv.org/abs/2012.04781\n\nHistorically, additionally employing the VGG-based perceptual loss has helped to overcome this issue, significantly improving the synthesis quality, but at the same time limiting the progress of GAN models for semantic image synthesis.\n\nIn this work, they propose a novel, simplified GAN model, which needs only adversarial supervision to achieve high quality results.\n\nThey re-design the discriminator as a semantic segmentation network, directly using the given semantic label maps as the ground truth for training.\n\nBy providing stronger supervision to the discriminator as well as to the generator through spatially- and semantically-aware discriminator feedback, they are able to synthesize images of higher fidelity with better alignment to their input label maps, making the use of the perceptual loss superfluous.\n\n**Authors -> \nVadim Sushko, Edgar Schoenfeld, Dan Zhang, Juergen Gall, Bernt Schiele, Anna Khoreva**", "1157002": "Thanks for sharing @mobassir" }, "source": "meta" }