{
  "id": 112989,
  "title": "Best strategy for re-labeling train",
  "url": "/competitions/open-images-2019-instance-segmentation/discussion/112989",
  "author_name": "n01z3",
  "post_date": "2019-10-16T11:55:21.379000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I believe pre-trained networks are very valuable. I going to train compact networks for my work tasks. At the same time, training on the original annotation seems very unreasonable due to large number of omissions. But, I have rather good ensemble. However, it seemed to me that the mAP metric is very sensitive to false positives. Therefore, just cooked masks from ensemble taking all prediction above some threshold will most likely not work.\nIn this competition, I didn't pseudo-lableing and I don’t have such an intuition in detection tasks.\nIf anyone has such an experience, can you suggest an approach please?</p>",
  "messages": [
    {
      "id": 650448,
      "postDate": "2019-10-16T11:55:21.380Z",
      "content": "<p>I believe pre-trained networks are very valuable. I going to train compact networks for my work tasks. At the same time, training on the original annotation seems very unreasonable due to large number of omissions. But, I have rather good ensemble. However, it seemed to me that the mAP metric is very sensitive to false positives. Therefore, just cooked masks from ensemble taking all prediction above some threshold will most likely not work.\nIn this competition, I didn't pseudo-lableing and I don’t have such an intuition in detection tasks.\nIf anyone has such an experience, can you suggest an approach please?</p>",
      "rawMarkdown": "I believe pre-trained networks are very valuable. I going to train compact networks for my work tasks. At the same time, training on the original annotation seems very unreasonable due to large number of omissions. But, I have rather good ensemble. However, it seemed to me that the mAP metric is very sensitive to false positives. Therefore, just cooked masks from ensemble taking all prediction above some threshold will most likely not work.\nIn this competition, I didn't pseudo-lableing and I don’t have such an intuition in detection tasks.\nIf anyone has such an experience, can you suggest an approach please?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "650448": "I believe pre-trained networks are very valuable. I going to train compact networks for my work tasks. At the same time, training on the original annotation seems very unreasonable due to large number of omissions. But, I have rather good ensemble. However, it seemed to me that the mAP metric is very sensitive to false positives. Therefore, just cooked masks from ensemble taking all prediction above some threshold will most likely not work.\nIn this competition, I didn't pseudo-lableing and I don’t have such an intuition in detection tasks.\nIf anyone has such an experience, can you suggest an approach please?"
  }
}