{
  "id": 420120,
  "title": "Please help! The Usage of TTA in Instance Segmentation",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/420120",
  "author_name": "",
  "post_date": "2023-06-29T09:35:59.547494200Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hello, I would like to verify if TTA has improved my model, but I do not understand how TTA should be used in instance segmentation to obtain the final result. My current approach is to invert the mask obtained after flipping and rotating onto the original image, but the number of instance segmentation objects they obtain is inconsistent, which makes me very confused</p>",
  "messages": [
    {
      "id": "2322446",
      "postDate": "06/29/2023 09:35:59",
      "content": "<p>Hello, I would like to verify if TTA has improved my model, but I do not understand how TTA should be used in instance segmentation to obtain the final result. My current approach is to invert the mask obtained after flipping and rotating onto the original image, but the number of instance segmentation objects they obtain is inconsistent, which makes me very confused</p>",
      "rawMarkdown": "Hello, I would like to verify if TTA has improved my model, but I do not understand how TTA should be used in instance segmentation to obtain the final result. My current approach is to invert the mask obtained after flipping and rotating onto the original image, but the number of instance segmentation objects they obtain is inconsistent, which makes me very confused",
      "votes": null
    },
    {
      "id": "2322695",
      "postDate": "06/29/2023 12:47:13",
      "content": "<p>For TTA<br>\nYou transform the images , in ways you think would help your results<br>\nI'd suggest that limit the transforms to geometric ones. Like rotate (in multiples   of 90), zoom, affine<br>\nAnd then you can perform inverse transformation accordingly</p>\n<p>Reason you're getting different number of Segmentation objects is you're probably distorting the images like elastic transform</p>",
      "rawMarkdown": "For TTA\nYou transform the images , in ways you think would help your results\nI'd suggest that limit the transforms to geometric ones. Like rotate (in multiples   of 90), zoom, affine\nAnd then you can perform inverse transformation accordingly\n\n\nReason you're getting different number of Segmentation objects is you're probably distorting the images like elastic transform",
      "votes": null
    },
    {
      "id": "2329226",
      "postDate": "07/04/2023 06:47:20",
      "content": "<p>I also used TTA, and when flipped horizontally, the number of segments is the same as the original image, but the number of segments flipped vertically is different. What is the reason for this?</p>",
      "rawMarkdown": "I also used TTA, and when flipped horizontally, the number of segments is the same as the original image, but the number of segments flipped vertically is different. What is the reason for this?",
      "votes": null
    },
    {
      "id": "2361386",
      "postDate": "07/27/2023 11:24:15",
      "content": "<p>Hey could you share what TTA proved most beneficial, for me vertical and horizontal flips just decrease the score, while rotation led to some improvements</p>",
      "rawMarkdown": "Hey could you share what TTA proved most beneficial, for me vertical and horizontal flips just decrease the score, while rotation led to some improvements",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2322695,
      "author_name": "bhavesjain",
      "author_url": "",
      "post_date": "06/29/2023 12:47:13",
      "content": "<p>For TTA<br>\nYou transform the images , in ways you think would help your results<br>\nI'd suggest that limit the transforms to geometric ones. Like rotate (in multiples   of 90), zoom, affine<br>\nAnd then you can perform inverse transformation accordingly</p>\n<p>Reason you're getting different number of Segmentation objects is you're probably distorting the images like elastic transform</p>",
      "votes": null,
      "replies": [
        {
          "id": 2329226,
          "author_name": "fanshutou",
          "author_url": "",
          "post_date": "07/04/2023 06:47:20",
          "content": "<p>I also used TTA, and when flipped horizontally, the number of segments is the same as the original image, but the number of segments flipped vertically is different. What is the reason for this?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2361386,
      "author_name": "josipvrdoljak",
      "author_url": "",
      "post_date": "07/27/2023 11:24:15",
      "content": "<p>Hey could you share what TTA proved most beneficial, for me vertical and horizontal flips just decrease the score, while rotation led to some improvements</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2322446": "Hello, I would like to verify if TTA has improved my model, but I do not understand how TTA should be used in instance segmentation to obtain the final result. My current approach is to invert the mask obtained after flipping and rotating onto the original image, but the number of instance segmentation objects they obtain is inconsistent, which makes me very confused",
    "2322695": "For TTA\nYou transform the images , in ways you think would help your results\nI'd suggest that limit the transforms to geometric ones. Like rotate (in multiples   of 90), zoom, affine\nAnd then you can perform inverse transformation accordingly\n\n\nReason you're getting different number of Segmentation objects is you're probably distorting the images like elastic transform",
    "2329226": "I also used TTA, and when flipped horizontally, the number of segments is the same as the original image, but the number of segments flipped vertically is different. What is the reason for this?",
    "2361386": "Hey could you share what TTA proved most beneficial, for me vertical and horizontal flips just decrease the score, while rotation led to some improvements"
  },
  "source": "meta"
}