{ "id": 303791, "title": "Style Transfer - Pytorch", "url": "/competitions/tensorflow-great-barrier-reef/discussion/303791", "author_name": "", "post_date": "2022-01-29T12:27:48.127149100Z", "votes": 10, "comment_count": 6, "views": 0, "content": "

Hello :)

\n

Please check the style transfer notebook on how to train the model and make NST (Neural Style Transfer)

\n

How to make pictures from the internet looks like from the dataset.
\nhttps://www.kaggle.com/marcinstasko/cots-neuralstyle-transfer-pytorch-augumentation

", "messages": [ { "id": "1668187", "postDate": "01/29/2022 12:27:48", "content": "

Hello :)

\n

Please check the style transfer notebook on how to train the model and make NST (Neural Style Transfer)

\n

How to make pictures from the internet looks like from the dataset.
\nhttps://www.kaggle.com/marcinstasko/cots-neuralstyle-transfer-pytorch-augumentation

", "rawMarkdown": "Hello :) \n\nPlease check the style transfer notebook on how to train the model and make NST (Neural Style Transfer)\n\nHow to make pictures from the internet looks like from the dataset.\nhttps://www.kaggle.com/marcinstasko/cots-neuralstyle-transfer-pytorch-augumentation", "votes": null }, { "id": "1668262", "postDate": "01/29/2022 13:41:29", "content": "

Shall unsupervised learning methods like contrast method or transfer learning with unsupervised task used by top solution in this competition ?

\n

I think this competition is hard to use zoom in methods like SNIPER
\nbecause as my test result it's hard to find patches to zoom in
\ndue to there background reefs it very similar to cots make it hard to figure out .
\neven use focal loss or carefully debug , recall still keeping poor
\nmake this advantage turn out to be a disadvantage bottonneck

\n

Generate training sample is also hardly to effect
\nbecause model always have good ability to figure out \"noise\" or evidences of \"blending\"
\nas experiences in past competition

\n

This competition is really hard
\ntarget is sparse , positive negative is imbalance. not so many information can be used (video frames continuity?)
\nup to now only priciple idea make sense is bigger model or higher resolution,

\n

Maybe only ways to go on 0.7 is \"innovation\" with something deeply understanding as no others known
\nbecause I search for so many sota methods and recently invest resouces , I cannot find something really make sense

", "rawMarkdown": "Shall unsupervised learning methods like contrast method or transfer learning with unsupervised task used by top solution in this competition ?\n\nI think this competition is hard to use zoom in methods like SNIPER \nbecause as my test result it's hard to find patches to zoom in\ndue to there background reefs it very similar to cots make it hard to figure out .\neven use focal loss or carefully debug , recall still keeping poor\nmake this advantage turn out to be a disadvantage bottonneck\n\nGenerate training sample is also hardly to effect \nbecause model always have good ability to figure out \"noise\" or evidences of \"blending\"\nas experiences in past competition \n\nThis competition is really hard \ntarget is sparse , positive negative is imbalance. not so many information can be used (video frames continuity?)\nup to now only priciple idea make sense is bigger model or higher resolution,\n\nMaybe only ways to go on 0.7 is \"innovation\" with something deeply understanding as no others known\nbecause I search for so many sota methods and recently invest resouces , I cannot find something really make sense", "votes": null }, { "id": "1668270", "postDate": "01/29/2022 13:42:22", "content": "

🤔

", "rawMarkdown": "🤔", "votes": null }, { "id": "1668278", "postDate": "01/29/2022 13:50:23", "content": "

Unfortunately, Style Transfer or just blending are problematic because we need to carefully catch the starfish not to blend starfish with the sea. They are separate objects in the real pictures.

", "rawMarkdown": "Unfortunately, Style Transfer or just blending are problematic because we need to carefully catch the starfish not to blend starfish with the sea. They are separate objects in the real pictures.", "votes": null }, { "id": "1670467", "postDate": "01/31/2022 14:51:33", "content": "

You can check another augmentation method using GAN
\nGenerate synthetical COTS from random noise:
\nNotebook: link
\nDiscussion: link

", "rawMarkdown": "You can check another augmentation method using GAN\nGenerate synthetical COTS from random noise:\nNotebook: [link](https://www.kaggle.com/marcinstasko/unlimited-cots-generator-pytorch-gan-in-action)\nDiscussion: [link](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/303988)", "votes": null }, { "id": "1670527", "postDate": "01/31/2022 15:57:59", "content": "

you can:

\n
    \n
  1. sample some images
    2. \n
  2. open up photoshop, draw filled circles on the reef. the filled circles represent the size of cots.
    3. \n
  3. build a predictor to predict the size of starfish on reef. use this size to synthesize/blend images
    4. \n
\n

this can be finished in one day

\n
\n

also:

\n
    \n
  1. sample same image.
    2. \n
  2. label two class: reef and non reef(e.g sea region)
    3. \n
  3. train a semantic segmentation net.
    4. \n
\n
\n

you need not many images for these. also this additional models can help you in your prediction too

", "rawMarkdown": "you can:\n1. sample some images\n2. open up photoshop, draw filled circles on the reef. the filled circles represent the size of cots.\n3. build a predictor to predict the size of starfish on reef. use this size to synthesize/blend images\n\nthis can be finished in one day\n\n----\n\nalso:\n1. sample same image.\n2. label two class: reef and non reef(e.g sea region)\n3. train a semantic segmentation net.\n\n----\n\nyou need not many images for these. also this additional models can help you in your prediction too", "votes": null }, { "id": "1670559", "postDate": "01/31/2022 16:26:29", "content": "

Thank you @hengck23 for your comment. I have not tried SS yet. I have struggled so much to detect the SF and background 😂

", "rawMarkdown": "Thank you @hengck23 for your comment. I have not tried SS yet. I have struggled so much to detect the SF and background 😂", "votes": null } ], "comments": [ { "id": 1668262, "author_name": "drzhuzhe", "author_url": "", "post_date": "01/29/2022 13:41:29", "content": "

Shall unsupervised learning methods like contrast method or transfer learning with unsupervised task used by top solution in this competition ?

\n

I think this competition is hard to use zoom in methods like SNIPER
\nbecause as my test result it's hard to find patches to zoom in
\ndue to there background reefs it very similar to cots make it hard to figure out .
\neven use focal loss or carefully debug , recall still keeping poor
\nmake this advantage turn out to be a disadvantage bottonneck

\n

Generate training sample is also hardly to effect
\nbecause model always have good ability to figure out \"noise\" or evidences of \"blending\"
\nas experiences in past competition

\n

This competition is really hard
\ntarget is sparse , positive negative is imbalance. not so many information can be used (video frames continuity?)
\nup to now only priciple idea make sense is bigger model or higher resolution,

\n

Maybe only ways to go on 0.7 is \"innovation\" with something deeply understanding as no others known
\nbecause I search for so many sota methods and recently invest resouces , I cannot find something really make sense

", "votes": null, "replies": [ { "id": 1668278, "author_name": "marcinstasko", "author_url": "", "post_date": "01/29/2022 13:50:23", "content": "

Unfortunately, Style Transfer or just blending are problematic because we need to carefully catch the starfish not to blend starfish with the sea. They are separate objects in the real pictures.

", "votes": null, "replies": [] }, { "id": 1670527, "author_name": "hengck23", "author_url": "", "post_date": "01/31/2022 15:57:59", "content": "

you can:

\n
    \n
  1. sample some images
    2. \n
  2. open up photoshop, draw filled circles on the reef. the filled circles represent the size of cots.
    3. \n
  3. build a predictor to predict the size of starfish on reef. use this size to synthesize/blend images
    4. \n
\n

this can be finished in one day

\n
\n

also:

\n
    \n
  1. sample same image.
    2. \n
  2. label two class: reef and non reef(e.g sea region)
    3. \n
  3. train a semantic segmentation net.
    4. \n
\n
\n

you need not many images for these. also this additional models can help you in your prediction too

", "votes": null, "replies": [] }, { "id": 1670559, "author_name": "marcinstasko", "author_url": "", "post_date": "01/31/2022 16:26:29", "content": "

Thank you @hengck23 for your comment. I have not tried SS yet. I have struggled so much to detect the SF and background 😂

", "votes": null, "replies": [] } ] }, { "id": 1668270, "author_name": "robsonsan", "author_url": "", "post_date": "01/29/2022 13:42:22", "content": "

🤔

", "votes": null, "replies": [] }, { "id": 1670467, "author_name": "marcinstasko", "author_url": "", "post_date": "01/31/2022 14:51:33", "content": "

You can check another augmentation method using GAN
\nGenerate synthetical COTS from random noise:
\nNotebook: link
\nDiscussion: link

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1668187": "Hello :) \n\nPlease check the style transfer notebook on how to train the model and make NST (Neural Style Transfer)\n\nHow to make pictures from the internet looks like from the dataset.\nhttps://www.kaggle.com/marcinstasko/cots-neuralstyle-transfer-pytorch-augumentation", "1668262": "Shall unsupervised learning methods like contrast method or transfer learning with unsupervised task used by top solution in this competition ?\n\nI think this competition is hard to use zoom in methods like SNIPER \nbecause as my test result it's hard to find patches to zoom in\ndue to there background reefs it very similar to cots make it hard to figure out .\neven use focal loss or carefully debug , recall still keeping poor\nmake this advantage turn out to be a disadvantage bottonneck\n\nGenerate training sample is also hardly to effect \nbecause model always have good ability to figure out \"noise\" or evidences of \"blending\"\nas experiences in past competition \n\nThis competition is really hard \ntarget is sparse , positive negative is imbalance. not so many information can be used (video frames continuity?)\nup to now only priciple idea make sense is bigger model or higher resolution,\n\nMaybe only ways to go on 0.7 is \"innovation\" with something deeply understanding as no others known\nbecause I search for so many sota methods and recently invest resouces , I cannot find something really make sense", "1668270": "🤔", "1668278": "Unfortunately, Style Transfer or just blending are problematic because we need to carefully catch the starfish not to blend starfish with the sea. They are separate objects in the real pictures.", "1670467": "You can check another augmentation method using GAN\nGenerate synthetical COTS from random noise:\nNotebook: [link](https://www.kaggle.com/marcinstasko/unlimited-cots-generator-pytorch-gan-in-action)\nDiscussion: [link](https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/303988)", "1670527": "you can:\n1. sample some images\n2. open up photoshop, draw filled circles on the reef. the filled circles represent the size of cots.\n3. build a predictor to predict the size of starfish on reef. use this size to synthesize/blend images\n\nthis can be finished in one day\n\n----\n\nalso:\n1. sample same image.\n2. label two class: reef and non reef(e.g sea region)\n3. train a semantic segmentation net.\n\n----\n\nyou need not many images for these. also this additional models can help you in your prediction too", "1670559": "Thank you @hengck23 for your comment. I have not tried SS yet. I have struggled so much to detect the SF and background 😂" }, "source": "meta" }