{
  "id": 207912,
  "title": "Images with No Findings",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/207912",
  "author_name": "Yerram Varun",
  "post_date": "2020-12-31T22:32:58.025000",
  "votes": 3,
  "comment_count": 1,
  "views": 0,
  "content": "<p>There are a large number of images with no findings in the dataset. Currently we have notebooks which can predict bounding boxes for the 14 classes.</p>\n<p>YOLO v5 - <a href=\"https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class\" target=\"_blank\">https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class</a><br>\nFaster RCNN - <a href=\"https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class\" target=\"_blank\">https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class</a></p>\n<p>If we train models with the current data. The model will not perform well because the data is skewed.<br>\nWhat can we do to deal with this?<br>\nPossible Ideas -</p>\n<ol>\n<li>Oversampling / Undersampling</li>\n<li>Extensive Augmentation for regularization.</li>\n</ol>",
  "messages": [
    {
      "id": 1134147,
      "postDate": "2020-12-31T22:32:58.027Z",
      "content": "<p>There are a large number of images with no findings in the dataset. Currently we have notebooks which can predict bounding boxes for the 14 classes.</p>\n<p>YOLO v5 - <a href=\"https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class\" target=\"_blank\">https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class</a><br>\nFaster RCNN - <a href=\"https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class\" target=\"_blank\">https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class</a></p>\n<p>If we train models with the current data. The model will not perform well because the data is skewed.<br>\nWhat can we do to deal with this?<br>\nPossible Ideas -</p>\n<ol>\n<li>Oversampling / Undersampling</li>\n<li>Extensive Augmentation for regularization.</li>\n</ol>",
      "rawMarkdown": "There are a large number of images with no findings in the dataset. Currently we have notebooks which can predict bounding boxes for the 14 classes.\n \nYOLO v5 - https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class\nFaster RCNN - https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class\n\nIf we train models with the current data. The model will not perform well because the data is skewed.\nWhat can we do to deal with this?\nPossible Ideas -\n1. Oversampling / Undersampling\n2. Extensive Augmentation for regularization.",
      "votes": 3
    },
    {
      "id": 1137031,
      "postDate": "2021-01-03T16:03:36.267Z",
      "content": "<p>For me, oversampling seems to be helpful based on local CV</p>",
      "rawMarkdown": "For me, oversampling seems to be helpful based on local CV"
    }
  ],
  "comments": [
    {
      "id": 1137031,
      "author_name": "Sanyam Bhutani",
      "author_url": "",
      "post_date": "2021-01-03T16:03:36.267000",
      "content": "<p>For me, oversampling seems to be helpful based on local CV</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1134147": "There are a large number of images with no findings in the dataset. Currently we have notebooks which can predict bounding boxes for the 14 classes.\n \nYOLO v5 - https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class\nFaster RCNN - https://www.kaggle.com/yerramvarun/pytorch-fasterrcnn-with-group-kfold-14-class\n\nIf we train models with the current data. The model will not perform well because the data is skewed.\nWhat can we do to deal with this?\nPossible Ideas -\n1. Oversampling / Undersampling\n2. Extensive Augmentation for regularization.",
    "1137031": "For me, oversampling seems to be helpful based on local CV"
  }
}