{
  "id": 67679,
  "title": "Ensemble Approach For RSNA ",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/67679",
  "author_name": "",
  "post_date": "2018-10-04T12:37:07.643357200Z",
  "votes": 1,
  "comment_count": 3,
  "views": 0,
  "content": "<p>How can we ensemble outputs from different models? \nMy approach is to predict the bounding box only if all the models detected the positive case but, this is not working as expected!\nWhat can be the better way to do it in such cases?</p>",
  "messages": [
    {
      "id": "398662",
      "postDate": "10/04/2018 12:37:07",
      "content": "<p>How can we ensemble outputs from different models? \nMy approach is to predict the bounding box only if all the models detected the positive case but, this is not working as expected!\nWhat can be the better way to do it in such cases?</p>",
      "rawMarkdown": "How can we ensemble outputs from different models? \nMy approach is to predict the bounding box only if all the models detected the positive case but, this is not working as expected!\nWhat can be the better way to do it in such cases?",
      "votes": null
    },
    {
      "id": "400601",
      "postDate": "10/08/2018 16:06:11",
      "content": "<p>It would be possible to implement a non-max suppression along the lines of this work for nuclei detection (<a href=\"https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419\">https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419</a>)</p>",
      "rawMarkdown": "It would be possible to implement a non-max suppression along the lines of this work for nuclei detection (https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419)",
      "votes": null
    },
    {
      "id": "401048",
      "postDate": "10/09/2018 10:46:21",
      "content": "<p>Thanks Rob. I already tried this approach. I want to know if this approach is working for you because my score did not improve using this.  </p>",
      "rawMarkdown": "Thanks Rob. I already tried this approach. I want to know if this approach is working for you because my score did not improve using this.",
      "votes": null
    },
    {
      "id": "401140",
      "postDate": "10/09/2018 13:51:26",
      "content": "<p>Hi Adish :D\nEvery model will output segmentation mask .. this mask results when thresholding the output probability ..\nSo, every model will output a probability map .. just average these maps and perform thresholding on the average</p>",
      "rawMarkdown": "Hi Adish :D\nEvery model will output segmentation mask .. this mask results when thresholding the output probability ..\nSo, every model will output a probability map .. just average these maps and perform thresholding on the average",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 400601,
      "author_name": "robrasmussen",
      "author_url": "",
      "post_date": "10/08/2018 16:06:11",
      "content": "<p>It would be possible to implement a non-max suppression along the lines of this work for nuclei detection (<a href=\"https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419\">https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419</a>)</p>",
      "votes": null,
      "replies": [
        {
          "id": 401048,
          "author_name": "adish333",
          "author_url": "",
          "post_date": "10/09/2018 10:46:21",
          "content": "<p>Thanks Rob. I already tried this approach. I want to know if this approach is working for you because my score did not improve using this.  </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 401140,
      "author_name": "ahmedhshahin",
      "author_url": "",
      "post_date": "10/09/2018 13:51:26",
      "content": "<p>Hi Adish :D\nEvery model will output segmentation mask .. this mask results when thresholding the output probability ..\nSo, every model will output a probability map .. just average these maps and perform thresholding on the average</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "398662": "How can we ensemble outputs from different models? \nMy approach is to predict the bounding box only if all the models detected the positive case but, this is not working as expected!\nWhat can be the better way to do it in such cases?",
    "400601": "It would be possible to implement a non-max suppression along the lines of this work for nuclei detection (https://www.kaggle.com/mpware/ensembling-on-instance-segmentation-lb-0-419)",
    "401048": "Thanks Rob. I already tried this approach. I want to know if this approach is working for you because my score did not improve using this.",
    "401140": "Hi Adish :D\nEvery model will output segmentation mask .. this mask results when thresholding the output probability ..\nSo, every model will output a probability map .. just average these maps and perform thresholding on the average"
  },
  "source": "meta"
}