{
  "id": 469794,
  "title": "questions about 2.5d and axis(XY+XZ+YZ)",
  "url": "/competitions/blood-vessel-segmentation/discussion/469794",
  "author_name": "",
  "post_date": "2024-01-22T05:09:37.707248600Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>This is my first time participating in a segmentation competition, and I have learned a lot. But I still have two questions:</p>\n<ul>\n<li>About 2.5d method, how to determine the number of slices? I notice that some notebooks use 3 and 5, but I want to know if the larger the better without considering computing resources.</li>\n<li>About axis(XY+XZ+YZ), how much help can axis(XY+XZ+YZ) bring? The performance of my XZ and YZ models is not good.</li>\n</ul>\n<p>I hope someone can answer my doubts, which will deepen my understanding of 3D segmentation.🙂</p>",
  "messages": [
    {
      "id": "2613455",
      "postDate": "01/22/2024 05:09:37",
      "content": "<p>This is my first time participating in a segmentation competition, and I have learned a lot. But I still have two questions:</p>\n<ul>\n<li>About 2.5d method, how to determine the number of slices? I notice that some notebooks use 3 and 5, but I want to know if the larger the better without considering computing resources.</li>\n<li>About axis(XY+XZ+YZ), how much help can axis(XY+XZ+YZ) bring? The performance of my XZ and YZ models is not good.</li>\n</ul>\n<p>I hope someone can answer my doubts, which will deepen my understanding of 3D segmentation.🙂</p>",
      "rawMarkdown": "This is my first time participating in a segmentation competition, and I have learned a lot. But I still have two questions:\n- About 2.5d method, how to determine the number of slices? I notice that some notebooks use 3 and 5, but I want to know if the larger the better without considering computing resources.\n- About axis(XY+XZ+YZ), how much help can axis(XY+XZ+YZ) bring? The performance of my XZ and YZ models is not good.\n\nI hope someone can answer my doubts, which will deepen my understanding of 3D segmentation.🙂",
      "votes": null
    },
    {
      "id": "2613479",
      "postDate": "01/22/2024 05:17:58",
      "content": "<p>I have gotten up to 0.05 increase on LB because of it, but I have some models it actually hurts.</p>",
      "rawMarkdown": "I have gotten up to 0.05 increase on LB because of it, but I have some models it actually hurts.",
      "votes": null
    },
    {
      "id": "2613542",
      "postDate": "01/22/2024 05:53:26",
      "content": "<blockquote>\n  <p>The performance of my XZ and YZ models is not good.</p>\n</blockquote>\n<p>Why not make a model that does all of the axis ? at the end of the day it's job would be the same -&gt; segment the vasculature of a kidney.</p>\n<blockquote>\n  <p>How much help can axis(XY+XZ+YZ) bring?</p>\n</blockquote>\n<p>The real point of using the multiview is that we don't know along which axis the test set is going to be sliced. if your model is realy good at predicting along the z axis, and you feed it a y axis sliced input, you will get poor score (and maybe on private LB where you wouldn't even see it)</p>\n<blockquote>\n  <p>how to determine the number of slices?</p>\n</blockquote>\n<p>Test a few values, there is never a straight up answer to this, this comes along Hyperparams optimisation. Use what works best.</p>",
      "rawMarkdown": "> The performance of my XZ and YZ models is not good.\n\nWhy not make a model that does all of the axis ? at the end of the day it's job would be the same -> segment the vasculature of a kidney.\n\n> How much help can axis(XY+XZ+YZ) bring?\n\nThe real point of using the multiview is that we don't know along which axis the test set is going to be sliced. if your model is realy good at predicting along the z axis, and you feed it a y axis sliced input, you will get poor score (and maybe on private LB where you wouldn't even see it)\n\n> how to determine the number of slices?\n\nTest a few values, there is never a straight up answer to this, this comes along Hyperparams optimisation. Use what works best.",
      "votes": null
    },
    {
      "id": "2615890",
      "postDate": "01/23/2024 11:38:01",
      "content": "<ul>\n<li>- Congrats on participating in your first segmentation competition! Regarding your questions, here are some insights:</li>\n</ul>\n<ol>\n<li><p>Determining the number of slices in 2.5D: The number of slices in the 2.5D method can vary depending on the dataset and the specific problem you are trying to solve. In general, including more slices can capture more contextual information, but it may also increase computational costs. It's recommended to experiment with different slice numbers (e.g., 3, 5) and evaluate their impact on performance to find the optimal balance between accuracy and efficiency.</p></li>\n<li><p>Axis(XY+XZ+YZ) in 3D segmentation: Combining the XY, XZ, and YZ axes provides a holistic view of the 3D data, capturing multiple perspectives and potentially improving segmentation performance. However, if your XZ and YZ models are not performing well, there could be several reasons such as the particular characteristics of the dataset or the model architecture. It's worth investigating and experimenting with different techniques, such as data augmentation, model modifications, or pre/post-processing steps to enhance your results.</p></li>\n</ol>\n<p>Hope this helps! Keep exploring and learning. Good luck with your segmentation journey! 🙂</p>",
      "rawMarkdown": "Congrats on participating in your first segmentation competition! Regarding your questions, here are some insights:\n\n1. Determining the number of slices in 2.5D: The number of slices in the 2.5D method can vary depending on the dataset and the specific problem you are trying to solve. In general, including more slices can capture more contextual information, but it may also increase computational costs. It's recommended to experiment with different slice numbers (e.g., 3, 5) and evaluate their impact on performance to find the optimal balance between accuracy and efficiency.\n\n2. Axis(XY+XZ+YZ) in 3D segmentation: Combining the XY, XZ, and YZ axes provides a holistic view of the 3D data, capturing multiple perspectives and potentially improving segmentation performance. However, if your XZ and YZ models are not performing well, there could be several reasons such as the particular characteristics of the dataset or the model architecture. It's worth investigating and experimenting with different techniques, such as data augmentation, model modifications, or pre/post-processing steps to enhance your results.\n\nHope this helps! Keep exploring and learning. Good luck with your segmentation journey! 🙂",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2613479,
      "author_name": "chemdatafarmer",
      "author_url": "",
      "post_date": "01/22/2024 05:17:58",
      "content": "<p>I have gotten up to 0.05 increase on LB because of it, but I have some models it actually hurts.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2613542,
      "author_name": "janmpia",
      "author_url": "",
      "post_date": "01/22/2024 05:53:26",
      "content": "<blockquote>\n  <p>The performance of my XZ and YZ models is not good.</p>\n</blockquote>\n<p>Why not make a model that does all of the axis ? at the end of the day it's job would be the same -&gt; segment the vasculature of a kidney.</p>\n<blockquote>\n  <p>How much help can axis(XY+XZ+YZ) bring?</p>\n</blockquote>\n<p>The real point of using the multiview is that we don't know along which axis the test set is going to be sliced. if your model is realy good at predicting along the z axis, and you feed it a y axis sliced input, you will get poor score (and maybe on private LB where you wouldn't even see it)</p>\n<blockquote>\n  <p>how to determine the number of slices?</p>\n</blockquote>\n<p>Test a few values, there is never a straight up answer to this, this comes along Hyperparams optimisation. Use what works best.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2615890,
      "author_name": "bjarkason",
      "author_url": "",
      "post_date": "01/23/2024 11:38:01",
      "content": "<ul>\n<li>- Congrats on participating in your first segmentation competition! Regarding your questions, here are some insights:</li>\n</ul>\n<ol>\n<li><p>Determining the number of slices in 2.5D: The number of slices in the 2.5D method can vary depending on the dataset and the specific problem you are trying to solve. In general, including more slices can capture more contextual information, but it may also increase computational costs. It's recommended to experiment with different slice numbers (e.g., 3, 5) and evaluate their impact on performance to find the optimal balance between accuracy and efficiency.</p></li>\n<li><p>Axis(XY+XZ+YZ) in 3D segmentation: Combining the XY, XZ, and YZ axes provides a holistic view of the 3D data, capturing multiple perspectives and potentially improving segmentation performance. However, if your XZ and YZ models are not performing well, there could be several reasons such as the particular characteristics of the dataset or the model architecture. It's worth investigating and experimenting with different techniques, such as data augmentation, model modifications, or pre/post-processing steps to enhance your results.</p></li>\n</ol>\n<p>Hope this helps! Keep exploring and learning. Good luck with your segmentation journey! 🙂</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2613455": "This is my first time participating in a segmentation competition, and I have learned a lot. But I still have two questions:\n- About 2.5d method, how to determine the number of slices? I notice that some notebooks use 3 and 5, but I want to know if the larger the better without considering computing resources.\n- About axis(XY+XZ+YZ), how much help can axis(XY+XZ+YZ) bring? The performance of my XZ and YZ models is not good.\n\nI hope someone can answer my doubts, which will deepen my understanding of 3D segmentation.🙂",
    "2613479": "I have gotten up to 0.05 increase on LB because of it, but I have some models it actually hurts.",
    "2613542": "> The performance of my XZ and YZ models is not good.\n\nWhy not make a model that does all of the axis ? at the end of the day it's job would be the same -> segment the vasculature of a kidney.\n\n> How much help can axis(XY+XZ+YZ) bring?\n\nThe real point of using the multiview is that we don't know along which axis the test set is going to be sliced. if your model is realy good at predicting along the z axis, and you feed it a y axis sliced input, you will get poor score (and maybe on private LB where you wouldn't even see it)\n\n> how to determine the number of slices?\n\nTest a few values, there is never a straight up answer to this, this comes along Hyperparams optimisation. Use what works best.",
    "2615890": "Congrats on participating in your first segmentation competition! Regarding your questions, here are some insights:\n\n1. Determining the number of slices in 2.5D: The number of slices in the 2.5D method can vary depending on the dataset and the specific problem you are trying to solve. In general, including more slices can capture more contextual information, but it may also increase computational costs. It's recommended to experiment with different slice numbers (e.g., 3, 5) and evaluate their impact on performance to find the optimal balance between accuracy and efficiency.\n\n2. Axis(XY+XZ+YZ) in 3D segmentation: Combining the XY, XZ, and YZ axes provides a holistic view of the 3D data, capturing multiple perspectives and potentially improving segmentation performance. However, if your XZ and YZ models are not performing well, there could be several reasons such as the particular characteristics of the dataset or the model architecture. It's worth investigating and experimenting with different techniques, such as data augmentation, model modifications, or pre/post-processing steps to enhance your results.\n\nHope this helps! Keep exploring and learning. Good luck with your segmentation journey! 🙂"
  },
  "source": "meta"
}