{
  "id": 231102,
  "title": "Level of blur",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/231102",
  "author_name": "",
  "post_date": "2021-04-06T23:43:17.889549100Z",
  "votes": 2,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hi,</p>\n<p>I hope to learn if there is a way to measure the blurry-ness of test images. As you know some public test images are a bit blurry, and I think I've read someone saying that \"it was part of the task\". </p>\n<p>So an idea is to train some models with blurred images and some with clear images, and predict accordingly?</p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": "1265482",
      "postDate": "04/06/2021 23:43:17",
      "content": "<p>Hi,</p>\n<p>I hope to learn if there is a way to measure the blurry-ness of test images. As you know some public test images are a bit blurry, and I think I've read someone saying that \"it was part of the task\". </p>\n<p>So an idea is to train some models with blurred images and some with clear images, and predict accordingly?</p>\n<p>Thank you!</p>",
      "rawMarkdown": "Hi,\n\nI hope to learn if there is a way to measure the blurry-ness of test images. As you know some public test images are a bit blurry, and I think I've read someone saying that \"it was part of the task\". \n\nSo an idea is to train some models with blurred images and some with clear images, and predict accordingly?\n\nThank you!",
      "votes": null
    },
    {
      "id": "1268435",
      "postDate": "04/09/2021 11:50:10",
      "content": "<p>Use gaussian or similar blur as augmentation</p>",
      "rawMarkdown": "Use gaussian or similar blur as augmentation",
      "votes": null
    },
    {
      "id": "1270108",
      "postDate": "04/11/2021 09:50:20",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/reaverlee\" target=\"_blank\">@reaverlee</a> You can also try to add small randomness for things like: brightness, contrast, saturation, hue etc.<br>\nAs long as you keep the randomness for those properties small it could benefit your model training.</p>\n<p>If you set the allowed range for random values to large it could 'kill' your model performance. Brightness is for sure one you don't want to set to large. </p>",
      "rawMarkdown": "Hi @reaverlee You can also try to add small randomness for things like: brightness, contrast, saturation, hue etc.\nAs long as you keep the randomness for those properties small it could benefit your model training.\n\nIf you set the allowed range for random values to large it could 'kill' your model performance. Brightness is for sure one you don't want to set to large.",
      "votes": null
    },
    {
      "id": "1271414",
      "postDate": "04/12/2021 15:31:18",
      "content": "<p>Thank you both so much!!</p>",
      "rawMarkdown": "Thank you both so much!!",
      "votes": null
    },
    {
      "id": "1272401",
      "postDate": "04/13/2021 12:54:14",
      "content": "<p>You could train a model that takes blurry images as input and outputs non blurry images. (Train the model by using non blurry images as target and blurring the input).</p>",
      "rawMarkdown": "You could train a model that takes blurry images as input and outputs non blurry images. (Train the model by using non blurry images as target and blurring the input).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1268435,
      "author_name": "sakvaua",
      "author_url": "",
      "post_date": "04/09/2021 11:50:10",
      "content": "<p>Use gaussian or similar blur as augmentation</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1270108,
      "author_name": "rsmits",
      "author_url": "",
      "post_date": "04/11/2021 09:50:20",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/reaverlee\" target=\"_blank\">@reaverlee</a> You can also try to add small randomness for things like: brightness, contrast, saturation, hue etc.<br>\nAs long as you keep the randomness for those properties small it could benefit your model training.</p>\n<p>If you set the allowed range for random values to large it could 'kill' your model performance. Brightness is for sure one you don't want to set to large. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1271414,
      "author_name": "reaverlee",
      "author_url": "",
      "post_date": "04/12/2021 15:31:18",
      "content": "<p>Thank you both so much!!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1272401,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "04/13/2021 12:54:14",
      "content": "<p>You could train a model that takes blurry images as input and outputs non blurry images. (Train the model by using non blurry images as target and blurring the input).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1265482": "Hi,\n\nI hope to learn if there is a way to measure the blurry-ness of test images. As you know some public test images are a bit blurry, and I think I've read someone saying that \"it was part of the task\". \n\nSo an idea is to train some models with blurred images and some with clear images, and predict accordingly?\n\nThank you!",
    "1268435": "Use gaussian or similar blur as augmentation",
    "1270108": "Hi @reaverlee You can also try to add small randomness for things like: brightness, contrast, saturation, hue etc.\nAs long as you keep the randomness for those properties small it could benefit your model training.\n\nIf you set the allowed range for random values to large it could 'kill' your model performance. Brightness is for sure one you don't want to set to large.",
    "1271414": "Thank you both so much!!",
    "1272401": "You could train a model that takes blurry images as input and outputs non blurry images. (Train the model by using non blurry images as target and blurring the input)."
  },
  "source": "meta"
}