{
  "id": 418413,
  "title": "The problems encountered in completing this task using semantic segmentation models",
  "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/418413",
  "author_name": "bent1e",
  "post_date": "2023-06-20T13:51:14.676000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<ol>\n<li>Obtaining instance segmentation regions: I used skimage.measure.label to extract and filter the mask's regions (filtering out some connected regions with small areas)</li>\n<li>Obtaining confidence scores: I use the normalized average pixel values within the segmented region as the confidence score, and choose (&gt;0.3) as the final confidence score</li>\n<li>No sorting required: Recent experiments have shown that sorting or not has no impact on score<br>\nCurrent issue: If using semantic segmentation models, obtaining confidence scores is very important, and currently there is no good method to calculate confidence scores</li>\n</ol>",
  "messages": [
    {
      "id": 2310594,
      "postDate": "2023-06-20T13:51:14.677Z",
      "content": "<ol>\n<li>Obtaining instance segmentation regions: I used skimage.measure.label to extract and filter the mask's regions (filtering out some connected regions with small areas)</li>\n<li>Obtaining confidence scores: I use the normalized average pixel values within the segmented region as the confidence score, and choose (&gt;0.3) as the final confidence score</li>\n<li>No sorting required: Recent experiments have shown that sorting or not has no impact on score<br>\nCurrent issue: If using semantic segmentation models, obtaining confidence scores is very important, and currently there is no good method to calculate confidence scores</li>\n</ol>",
      "rawMarkdown": "1. Obtaining instance segmentation regions: I used skimage.measure.label to extract and filter the mask's regions (filtering out some connected regions with small areas)\n2. Obtaining confidence scores: I use the normalized average pixel values within the segmented region as the confidence score, and choose (>0.3) as the final confidence score\n3. No sorting required: Recent experiments have shown that sorting or not has no impact on score\nCurrent issue: If using semantic segmentation models, obtaining confidence scores is very important, and currently there is no good method to calculate confidence scores",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "2310594": "1. Obtaining instance segmentation regions: I used skimage.measure.label to extract and filter the mask's regions (filtering out some connected regions with small areas)\n2. Obtaining confidence scores: I use the normalized average pixel values within the segmented region as the confidence score, and choose (>0.3) as the final confidence score\n3. No sorting required: Recent experiments have shown that sorting or not has no impact on score\nCurrent issue: If using semantic segmentation models, obtaining confidence scores is very important, and currently there is no good method to calculate confidence scores"
  }
}