{
  "id": 240307,
  "title": "Multiclass Vs Multilabel for image-case",
  "url": "/competitions/siim-covid19-detection/discussion/240307",
  "author_name": "",
  "post_date": "2021-05-19T07:40:39.063534700Z",
  "votes": 4,
  "comment_count": 7,
  "views": 0,
  "content": "<p>In <code>train_image_label.csv</code> for one image we only have one label from <code>\"negative\", \"typical\", \"indeterminate\", \"atypical\"</code>. So it doesn't seem like a <strong>multi-label</strong> problem. But according to competition site ,</p>\n<pre><code>For each test image, you will be predicting a bounding box and class for all finding\n</code></pre>\n<p>Also,</p>\n<pre><code>Images in the test set may contain more than one object. For each object in a given test image, you must predict a class ID of \"opacity\", a confidence score, and bounding box in format `xmin ymin xmax ymax`.\n</code></pre>\n<p>So, though for image-case we have multiple <strong>objects</strong> or <strong>bounding boxes</strong> with <strong>class ID</strong> <code>opacity</code>, we only have only <strong>1</strong> class[<code>\"negative\", \"typical\", \"indeterminate\", \"atypical\"</code>] per image?</p>",
  "messages": [
    {
      "id": "1314503",
      "postDate": "05/19/2021 07:40:39",
      "content": "<p>In <code>train_image_label.csv</code> for one image we only have one label from <code>\"negative\", \"typical\", \"indeterminate\", \"atypical\"</code>. So it doesn't seem like a <strong>multi-label</strong> problem. But according to competition site ,</p>\n<pre><code>For each test image, you will be predicting a bounding box and class for all finding\n</code></pre>\n<p>Also,</p>\n<pre><code>Images in the test set may contain more than one object. For each object in a given test image, you must predict a class ID of \"opacity\", a confidence score, and bounding box in format `xmin ymin xmax ymax`.\n</code></pre>\n<p>So, though for image-case we have multiple <strong>objects</strong> or <strong>bounding boxes</strong> with <strong>class ID</strong> <code>opacity</code>, we only have only <strong>1</strong> class[<code>\"negative\", \"typical\", \"indeterminate\", \"atypical\"</code>] per image?</p>",
      "rawMarkdown": "In `train_image_label.csv` for one image we only have one label from `\"negative\", \"typical\", \"indeterminate\", \"atypical\"`. So it doesn't seem like a **multi-label** problem. But according to competition site ,\n````\nFor each test image, you will be predicting a bounding box and class for all finding\n````\nAlso,\n````\nImages in the test set may contain more than one object. For each object in a given test image, you must predict a class ID of \"opacity\", a confidence score, and bounding box in format `xmin ymin xmax ymax`.\n````\nSo, though for image-case we have multiple **objects** or **bounding boxes** with **class ID** `opacity`, we only have only **1** class[`\"negative\", \"typical\", \"indeterminate\", \"atypical\"`] per image?",
      "votes": null
    },
    {
      "id": "1314668",
      "postDate": "05/19/2021 10:03:31",
      "content": "<p>We need to make predictions at both study-level and image-level. So, it is a multi-label classification at study-level and object detection at image-level.</p>\n<p>You can find more details in this <a href=\"https://www.kaggle.com/c/siim-covid19-detection/discussion/240329\" target=\"_blank\">thread</a>.</p>",
      "rawMarkdown": "We need to make predictions at both study-level and image-level. So, it is a multi-label classification at study-level and object detection at image-level.\n\nYou can find more details in this [thread](https://www.kaggle.com/c/siim-covid19-detection/discussion/240329).",
      "votes": null
    },
    {
      "id": "1314673",
      "postDate": "05/19/2021 10:05:44",
      "content": "<p>for each image, there is one label class not <strong>multilple</strong> . This what confused me. </p>",
      "rawMarkdown": "for each image, there is one label class not **multilple** . This what confused me.",
      "votes": null
    },
    {
      "id": "1314681",
      "postDate": "05/19/2021 10:11:36",
      "content": "<p>Yes, each image has only one label. A study has multiple images and hence multiple labels.</p>",
      "rawMarkdown": "Yes, each image has only one label. A study has multiple images and hence multiple labels.",
      "votes": null
    },
    {
      "id": "1315564",
      "postDate": "05/19/2021 23:24:03",
      "content": "<p>Now I tackle as multilabel although only one label per image in the train.<br>\nWill compare with multiclass score soon.</p>",
      "rawMarkdown": "Now I tackle as multilabel although only one label per image in the train.\nWill compare with multiclass score soon.",
      "votes": null
    },
    {
      "id": "1334472",
      "postDate": "06/03/2021 14:24:00",
      "content": "<p>Hi! Is the multiclass result better?</p>",
      "rawMarkdown": "Hi! Is the multiclass result better?",
      "votes": null
    },
    {
      "id": "1334671",
      "postDate": "06/03/2021 17:17:15",
      "content": "<p>I don't think there are many studies with multiple labels, thus I don't think using sigmoid instead of softmax would make such a great difference. What do you think?</p>",
      "rawMarkdown": "I don't think there are many studies with multiple labels, thus I don't think using sigmoid instead of softmax would make such a great difference. What do you think?",
      "votes": null
    },
    {
      "id": "1374508",
      "postDate": "07/03/2021 11:30:29",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> Which approach works good for you? Multi-label or multi-class? Looks like I am a bit late here.</p>",
      "rawMarkdown": "awsaf49 Which approach works good for you? Multi-label or multi-class? Looks like I am a bit late here.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1314668,
      "author_name": "hassiahk",
      "author_url": "",
      "post_date": "05/19/2021 10:03:31",
      "content": "<p>We need to make predictions at both study-level and image-level. So, it is a multi-label classification at study-level and object detection at image-level.</p>\n<p>You can find more details in this <a href=\"https://www.kaggle.com/c/siim-covid19-detection/discussion/240329\" target=\"_blank\">thread</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1314673,
          "author_name": "awsaf49",
          "author_url": "",
          "post_date": "05/19/2021 10:05:44",
          "content": "<p>for each image, there is one label class not <strong>multilple</strong> . This what confused me. </p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1314681,
          "author_name": "hassiahk",
          "author_url": "",
          "post_date": "05/19/2021 10:11:36",
          "content": "<p>Yes, each image has only one label. A study has multiple images and hence multiple labels.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1315564,
      "author_name": "drtausamaru",
      "author_url": "",
      "post_date": "05/19/2021 23:24:03",
      "content": "<p>Now I tackle as multilabel although only one label per image in the train.<br>\nWill compare with multiclass score soon.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1334472,
          "author_name": "klawensliu",
          "author_url": "",
          "post_date": "06/03/2021 14:24:00",
          "content": "<p>Hi! Is the multiclass result better?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1334671,
      "author_name": "josephamigo",
      "author_url": "",
      "post_date": "06/03/2021 17:17:15",
      "content": "<p>I don't think there are many studies with multiple labels, thus I don't think using sigmoid instead of softmax would make such a great difference. What do you think?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1374508,
      "author_name": "debarshichanda",
      "author_url": "",
      "post_date": "07/03/2021 11:30:29",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> Which approach works good for you? Multi-label or multi-class? Looks like I am a bit late here.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1314503": "In `train_image_label.csv` for one image we only have one label from `\"negative\", \"typical\", \"indeterminate\", \"atypical\"`. So it doesn't seem like a **multi-label** problem. But according to competition site ,\n````\nFor each test image, you will be predicting a bounding box and class for all finding\n````\nAlso,\n````\nImages in the test set may contain more than one object. For each object in a given test image, you must predict a class ID of \"opacity\", a confidence score, and bounding box in format `xmin ymin xmax ymax`.\n````\nSo, though for image-case we have multiple **objects** or **bounding boxes** with **class ID** `opacity`, we only have only **1** class[`\"negative\", \"typical\", \"indeterminate\", \"atypical\"`] per image?",
    "1314668": "We need to make predictions at both study-level and image-level. So, it is a multi-label classification at study-level and object detection at image-level.\n\nYou can find more details in this [thread](https://www.kaggle.com/c/siim-covid19-detection/discussion/240329).",
    "1314673": "for each image, there is one label class not **multilple** . This what confused me.",
    "1314681": "Yes, each image has only one label. A study has multiple images and hence multiple labels.",
    "1315564": "Now I tackle as multilabel although only one label per image in the train.\nWill compare with multiclass score soon.",
    "1334472": "Hi! Is the multiclass result better?",
    "1334671": "I don't think there are many studies with multiple labels, thus I don't think using sigmoid instead of softmax would make such a great difference. What do you think?",
    "1374508": "awsaf49 Which approach works good for you? Multi-label or multi-class? Looks like I am a bit late here."
  },
  "source": "meta"
}