{
  "id": 286159,
  "title": "Filtering predictions",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/286159",
  "author_name": "It's over 9,000!",
  "post_date": "2021-11-08T08:12:55.870000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Typical IOU vs predicted score of an image from mask rcnn:</p>\n<p><img src=\"https://ibb.co/1GbqyvG\"></p>\n<p>Anything prediction with IOU above 0.5 is good. As you can see, A low predicted score does not necessarily mean low IOU. There are consistently plenty of good predictions with low scores. The problem is we need ways to filter out the ones with low IOUs.</p>\n<p>Some ideas:</p>\n<ul>\n<li><strong>Minimum mask size filter</strong> (like in this <a href=\"https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference\" target=\"_blank\">notebook</a>): Since submission does not allow overlap, the predicted mask with lower score got truncated by ones with higher scores. </li>\n<li><strong>Separate disconnected masks to individual</strong>: Again an artifact of truncating lower score ones. Separating them out, each of them could more likely be filtered by the min size filter.</li>\n</ul>\n<p>Any other ideas?</p>",
  "messages": [
    {
      "id": 1575186,
      "postDate": "2021-11-08T08:12:55.870Z",
      "content": "<p>Typical IOU vs predicted score of an image from mask rcnn:</p>\n<p><img src=\"https://ibb.co/1GbqyvG\"></p>\n<p>Anything prediction with IOU above 0.5 is good. As you can see, A low predicted score does not necessarily mean low IOU. There are consistently plenty of good predictions with low scores. The problem is we need ways to filter out the ones with low IOUs.</p>\n<p>Some ideas:</p>\n<ul>\n<li><strong>Minimum mask size filter</strong> (like in this <a href=\"https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference\" target=\"_blank\">notebook</a>): Since submission does not allow overlap, the predicted mask with lower score got truncated by ones with higher scores. </li>\n<li><strong>Separate disconnected masks to individual</strong>: Again an artifact of truncating lower score ones. Separating them out, each of them could more likely be filtered by the min size filter.</li>\n</ul>\n<p>Any other ideas?</p>",
      "rawMarkdown": "Typical IOU vs predicted score of an image from mask rcnn:\n\n\n<img src=\"https://ibb.co/1GbqyvG\">\n\nAnything prediction with IOU above 0.5 is good. As you can see, A low predicted score does not necessarily mean low IOU. There are consistently plenty of good predictions with low scores. The problem is we need ways to filter out the ones with low IOUs.\n\nSome ideas:\n- **Minimum mask size filter** (like in this [notebook](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference)): Since submission does not allow overlap, the predicted mask with lower score got truncated by ones with higher scores. \n- **Separate disconnected masks to individual**: Again an artifact of truncating lower score ones. Separating them out, each of them could more likely be filtered by the min size filter.\n\nAny other ideas?\n",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1575186": "Typical IOU vs predicted score of an image from mask rcnn:\n\n\n<img src=\"https://ibb.co/1GbqyvG\">\n\nAnything prediction with IOU above 0.5 is good. As you can see, A low predicted score does not necessarily mean low IOU. There are consistently plenty of good predictions with low scores. The problem is we need ways to filter out the ones with low IOUs.\n\nSome ideas:\n- **Minimum mask size filter** (like in this [notebook](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference)): Since submission does not allow overlap, the predicted mask with lower score got truncated by ones with higher scores. \n- **Separate disconnected masks to individual**: Again an artifact of truncating lower score ones. Separating them out, each of them could more likely be filtered by the min size filter.\n\nAny other ideas?\n"
  }
}