Hi, I am new to semantic segmentation, I saw a lot of people use resnet as backbone of Unet, specifically, encoder of Unet, so I am wondering if the decoder structure of ResUnet is the same as the raw Unet https://arxiv.org/pdf/1505.04597.pdf, and if we can also change the decoder structure, which is better?
", "messages": [ { "id": "1123395", "postDate": "12/23/2020 07:26:58", "content": "Hi, I am new to semantic segmentation, I saw a lot of people use resnet as backbone of Unet, specifically, encoder of Unet, so I am wondering if the decoder structure of ResUnet is the same as the raw Unet https://arxiv.org/pdf/1505.04597.pdf, and if we can also change the decoder structure, which is better?
", "rawMarkdown": "Hi, I am new to semantic segmentation, I saw a lot of people use resnet as backbone of Unet, specifically, encoder of Unet, so I am wondering if the decoder structure of ResUnet is the same as the raw Unet https://arxiv.org/pdf/1505.04597.pdf, and if we can also change the decoder structure, which is better?", "votes": null }, { "id": "1126037", "postDate": "12/25/2020 09:31:09", "content": "Anything can be a decoder if you are brave enough. :)
\nResent-like decoder works fine.
thanks for replying, so you mean ResBolck can also be applied to upsample stage, and more often to achieve a decent result, at least no worse than raw upsample wo ResBlock, right?
", "rawMarkdown": "thanks for replying, so you mean ResBolck can also be applied to upsample stage, and more often to achieve a decent result, at least no worse than raw upsample wo ResBlock, right?", "votes": null }, { "id": "1126661", "postDate": "12/25/2020 19:10:58", "content": "There is no harm in using res-block for the decoder. There are variants of UNet in the literature including deep supervision, Res-UNet (both decoder and encoder), multi-input/multi-scale UNet. However, you should not forget that encoding the features are more important in a segmentation task, therefore, people mostly use a more robust encoder and a light-weight decoder to have a model with less trainable parameters. Increasing the depth and complexity of networks is a trad-off between improving accuracy and computational cost.
", "rawMarkdown": "There is no harm in using res-block for the decoder. There are variants of UNet in the literature including deep supervision, Res-UNet (both decoder and encoder), multi-input/multi-scale UNet. However, you should not forget that encoding the features are more important in a segmentation task, therefore, people mostly use a more robust encoder and a light-weight decoder to have a model with less trainable parameters. Increasing the depth and complexity of networks is a trad-off between improving accuracy and computational cost.", "votes": null }, { "id": "1126812", "postDate": "12/26/2020 01:38:35", "content": "Then all makes sense now. Thanks Aramis!
", "rawMarkdown": "Then all makes sense now. Thanks Aramis!", "votes": null } ], "comments": [ { "id": 1126037, "author_name": "sakvaua", "author_url": "", "post_date": "12/25/2020 09:31:09", "content": "Anything can be a decoder if you are brave enough. :)
\nResent-like decoder works fine.
thanks for replying, so you mean ResBolck can also be applied to upsample stage, and more often to achieve a decent result, at least no worse than raw upsample wo ResBlock, right?
", "votes": null, "replies": [] } ] }, { "id": 1126661, "author_name": "aramisvesal", "author_url": "", "post_date": "12/25/2020 19:10:58", "content": "There is no harm in using res-block for the decoder. There are variants of UNet in the literature including deep supervision, Res-UNet (both decoder and encoder), multi-input/multi-scale UNet. However, you should not forget that encoding the features are more important in a segmentation task, therefore, people mostly use a more robust encoder and a light-weight decoder to have a model with less trainable parameters. Increasing the depth and complexity of networks is a trad-off between improving accuracy and computational cost.
", "votes": null, "replies": [ { "id": 1126812, "author_name": "plugin1689", "author_url": "", "post_date": "12/26/2020 01:38:35", "content": "Then all makes sense now. Thanks Aramis!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1123395": "Hi, I am new to semantic segmentation, I saw a lot of people use resnet as backbone of Unet, specifically, encoder of Unet, so I am wondering if the decoder structure of ResUnet is the same as the raw Unet https://arxiv.org/pdf/1505.04597.pdf, and if we can also change the decoder structure, which is better?", "1126037": "Anything can be a decoder if you are brave enough. :)\nResent-like decoder works fine.", "1126046": "thanks for replying, so you mean ResBolck can also be applied to upsample stage, and more often to achieve a decent result, at least no worse than raw upsample wo ResBlock, right?", "1126661": "There is no harm in using res-block for the decoder. There are variants of UNet in the literature including deep supervision, Res-UNet (both decoder and encoder), multi-input/multi-scale UNet. However, you should not forget that encoding the features are more important in a segmentation task, therefore, people mostly use a more robust encoder and a light-weight decoder to have a model with less trainable parameters. Increasing the depth and complexity of networks is a trad-off between improving accuracy and computational cost.", "1126812": "Then all makes sense now. Thanks Aramis!" }, "source": "meta" }