{
  "id": 61906,
  "title": "How to fuse the results of multiple models？",
  "url": "/competitions/google-ai-open-images-object-detection-track/discussion/61906",
  "author_name": "reeds",
  "post_date": "2018-07-25T01:52:24.099000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I have test some state-of-the-art of object detection models on a challenge dataset. As shown in the table below. I have tried to fuse these results with nms(threshold=0.5) and then this simple method improves single model performance from 25.9% to 29.1%. Is there any other way besides this?</p>\n\n<p><img src=\"http://omyi7khf5.bkt.clouddn.com/QQ%E5%9B%BE%E7%89%8720180725094010.png\" alt=\"Results on val dataset\"></p>\n\n<p>I also recorded some of their other evaluation metrics (AP-small、AP-medium、AP-large).The greater the difference between these indicators between different models, the better the effect of fusion.</p>",
  "messages": [
    {
      "id": 361717,
      "postDate": "2018-07-25T01:52:24.100Z",
      "content": "<p>I have test some state-of-the-art of object detection models on a challenge dataset. As shown in the table below. I have tried to fuse these results with nms(threshold=0.5) and then this simple method improves single model performance from 25.9% to 29.1%. Is there any other way besides this?</p>\n\n<p><img src=\"http://omyi7khf5.bkt.clouddn.com/QQ%E5%9B%BE%E7%89%8720180725094010.png\" alt=\"Results on val dataset\"></p>\n\n<p>I also recorded some of their other evaluation metrics (AP-small、AP-medium、AP-large).The greater the difference between these indicators between different models, the better the effect of fusion.</p>",
      "rawMarkdown": "I have test some state-of-the-art of object detection models on a challenge dataset. As shown in the table below. I have tried to fuse these results with nms(threshold=0.5) and then this simple method improves single model performance from 25.9% to 29.1%. Is there any other way besides this?\n\n\n![Results on val dataset][1]\n\n\n  [1]: http://omyi7khf5.bkt.clouddn.com/QQ%E5%9B%BE%E7%89%8720180725094010.png\n\nI also recorded some of their other evaluation metrics (AP-small、AP-medium、AP-large).The greater the difference between these indicators between different models, the better the effect of fusion."
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "361717": "I have test some state-of-the-art of object detection models on a challenge dataset. As shown in the table below. I have tried to fuse these results with nms(threshold=0.5) and then this simple method improves single model performance from 25.9% to 29.1%. Is there any other way besides this?\n\n\n![Results on val dataset][1]\n\n\n  [1]: http://omyi7khf5.bkt.clouddn.com/QQ%E5%9B%BE%E7%89%8720180725094010.png\n\nI also recorded some of their other evaluation metrics (AP-small、AP-medium、AP-large).The greater the difference between these indicators between different models, the better the effect of fusion."
  }
}