{
  "id": 52250,
  "title": "anyone use google object detection framwork? what your LB score?",
  "url": "/competitions/data-science-bowl-2018/discussion/52250",
  "author_name": "",
  "post_date": "2018-03-18T00:43:38.785039800Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi guys,</p>\n\n<p>I use this object detection framework(mask faster rcnn) <a href=\"https://github.com/tensorflow/models/tree/master/research/object_detection\">object detection</a>\nI can only get about 0.32 LB score(not do any preprocess and postprocess). I have not see roi align in the code, not sure if this is the problem, this bad result will force me to change the network structure much.</p>\n\n<p>Anyone else use it? what your score?</p>\n\n<p>Thanks.</p>",
  "messages": [
    {
      "id": "297800",
      "postDate": "03/18/2018 00:43:38",
      "content": "<p>Hi guys,</p>\n\n<p>I use this object detection framework(mask faster rcnn) <a href=\"https://github.com/tensorflow/models/tree/master/research/object_detection\">object detection</a>\nI can only get about 0.32 LB score(not do any preprocess and postprocess). I have not see roi align in the code, not sure if this is the problem, this bad result will force me to change the network structure much.</p>\n\n<p>Anyone else use it? what your score?</p>\n\n<p>Thanks.</p>",
      "rawMarkdown": "Hi guys,\n\nI use this object detection framework(mask faster rcnn) [object detection][1]\nI can only get about 0.32 LB score(not do any preprocess and postprocess). I have not see roi align in the code, not sure if this is the problem, this bad result will force me to change the network structure much.\n\nAnyone else use it? what your score?\n\nThanks.\n\n  [1]: https://github.com/tensorflow/models/tree/master/research/object_detection",
      "votes": null
    },
    {
      "id": "297809",
      "postDate": "03/18/2018 01:34:16",
      "content": "<p>Although I do not use this code here, I like this code. I think roi align is not a problem. If I understand correctly, Detectron and tensorpack use real roi align. Maybe FPN is a problem for this tf det api. <br>\nBy the way,my wechat is wizard_123. I'd like to connect with more people in this field.</p>",
      "rawMarkdown": "Although I do not use this code here, I like this code. I think roi align is not a problem. If I understand correctly, Detectron and tensorpack use real roi align. Maybe FPN is a problem for this tf det api.    \nBy the way,my wechat is wizard_123. I'd like to connect with more people in this field.",
      "votes": null
    },
    {
      "id": "297874",
      "postDate": "03/18/2018 07:41:46",
      "content": "<p>Hi Yan,\nDoes your mask rcnn implemented FPN to get 0.4x?\nThanks</p>",
      "rawMarkdown": "Hi Yan,\nDoes your mask rcnn implemented FPN to get 0.4x?\nThanks",
      "votes": null
    },
    {
      "id": "297971",
      "postDate": "03/18/2018 12:47:58",
      "content": "<p>Both Matterport and Heng's code have FPN</p>",
      "rawMarkdown": "Both Matterport and Heng's code have FPN",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 297809,
      "author_name": "wayne999",
      "author_url": "",
      "post_date": "03/18/2018 01:34:16",
      "content": "<p>Although I do not use this code here, I like this code. I think roi align is not a problem. If I understand correctly, Detectron and tensorpack use real roi align. Maybe FPN is a problem for this tf det api. <br>\nBy the way,my wechat is wizard_123. I'd like to connect with more people in this field.</p>",
      "votes": null,
      "replies": [
        {
          "id": 297874,
          "author_name": "amxineohp",
          "author_url": "",
          "post_date": "03/18/2018 07:41:46",
          "content": "<p>Hi Yan,\nDoes your mask rcnn implemented FPN to get 0.4x?\nThanks</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 297971,
          "author_name": "wayne999",
          "author_url": "",
          "post_date": "03/18/2018 12:47:58",
          "content": "<p>Both Matterport and Heng's code have FPN</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "297800": "Hi guys,\n\nI use this object detection framework(mask faster rcnn) [object detection][1]\nI can only get about 0.32 LB score(not do any preprocess and postprocess). I have not see roi align in the code, not sure if this is the problem, this bad result will force me to change the network structure much.\n\nAnyone else use it? what your score?\n\nThanks.\n\n  [1]: https://github.com/tensorflow/models/tree/master/research/object_detection",
    "297809": "Although I do not use this code here, I like this code. I think roi align is not a problem. If I understand correctly, Detectron and tensorpack use real roi align. Maybe FPN is a problem for this tf det api.    \nBy the way,my wechat is wizard_123. I'd like to connect with more people in this field.",
    "297874": "Hi Yan,\nDoes your mask rcnn implemented FPN to get 0.4x?\nThanks",
    "297971": "Both Matterport and Heng's code have FPN"
  },
  "source": "meta"
}