{
  "id": 212967,
  "title": "Noisy Data (Possible mislabelings)",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/212967",
  "author_name": "",
  "post_date": "2021-01-20T22:46:30.721106300Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, I wanted to start a discussion regarding the noisy healthy leaves. The leaves seem to be mislabeled but may actually be labelled correctly, if this is part of the lifecycle.</p>\n<p>Do we know if the test data contains the same noisy healthy leaves? If so, what are some methods or ideas to help improve the model's accuracy on the test set?</p>",
  "messages": [
    {
      "id": "1162013",
      "postDate": "01/20/2021 22:46:30",
      "content": "<p>Hi, I wanted to start a discussion regarding the noisy healthy leaves. The leaves seem to be mislabeled but may actually be labelled correctly, if this is part of the lifecycle.</p>\n<p>Do we know if the test data contains the same noisy healthy leaves? If so, what are some methods or ideas to help improve the model's accuracy on the test set?</p>",
      "rawMarkdown": "Hi, I wanted to start a discussion regarding the noisy healthy leaves. The leaves seem to be mislabeled but may actually be labelled correctly, if this is part of the lifecycle.\n\nDo we know if the test data contains the same noisy healthy leaves? If so, what are some methods or ideas to help improve the model's accuracy on the test set?",
      "votes": null
    },
    {
      "id": "1162309",
      "postDate": "01/21/2021 05:14:20",
      "content": "<p>If I can find it in next few minutes I will post the link - but someone has posted a link to the paper written by the host for last years competition.  Combine that with the overview and it's clear to me that the hosts are aware of noise, that they tried some methods to reduce the noise in images when taking the photos, that some single plants can contain more than a single disease, etc.</p>\n<p>That leads me to the conclusion that they would like us to find a solution for noisy labels and therefore the test set will be just as noisy as the train.</p>\n<p>see this post to get to the paper - <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543</a></p>",
      "rawMarkdown": "If I can find it in next few minutes I will post the link - but someone has posted a link to the paper written by the host for last years competition.  Combine that with the overview and it's clear to me that the hosts are aware of noise, that they tried some methods to reduce the noise in images when taking the photos, that some single plants can contain more than a single disease, etc.\n\nThat leads me to the conclusion that they would like us to find a solution for noisy labels and therefore the test set will be just as noisy as the train.\n\nsee this post to get to the paper - https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1162309,
      "author_name": "pcjimmmy",
      "author_url": "",
      "post_date": "01/21/2021 05:14:20",
      "content": "<p>If I can find it in next few minutes I will post the link - but someone has posted a link to the paper written by the host for last years competition.  Combine that with the overview and it's clear to me that the hosts are aware of noise, that they tried some methods to reduce the noise in images when taking the photos, that some single plants can contain more than a single disease, etc.</p>\n<p>That leads me to the conclusion that they would like us to find a solution for noisy labels and therefore the test set will be just as noisy as the train.</p>\n<p>see this post to get to the paper - <a href=\"https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543\" target=\"_blank\">https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1162013": "Hi, I wanted to start a discussion regarding the noisy healthy leaves. The leaves seem to be mislabeled but may actually be labelled correctly, if this is part of the lifecycle.\n\nDo we know if the test data contains the same noisy healthy leaves? If so, what are some methods or ideas to help improve the model's accuracy on the test set?",
    "1162309": "If I can find it in next few minutes I will post the link - but someone has posted a link to the paper written by the host for last years competition.  Combine that with the overview and it's clear to me that the hosts are aware of noise, that they tried some methods to reduce the noise in images when taking the photos, that some single plants can contain more than a single disease, etc.\n\nThat leads me to the conclusion that they would like us to find a solution for noisy labels and therefore the test set will be just as noisy as the train.\n\nsee this post to get to the paper - https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/212543"
  },
  "source": "meta"
}