{
  "id": 94693,
  "title": "an immature idea",
  "url": "/competitions/open-images-2019-visual-relationship/discussion/94693",
  "author_name": "xbit",
  "post_date": "2019-06-06T09:18:17.562000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I plan to do it this way, if I have time in the next 3 months.</p>\n\n<p>Training:\n- Train a detection network with <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/data\">Detection Dataset</a>. After this, we can get the bbox easily.\n- Train a classification network with the <a href=\"https://www.kaggle.com/c/open-images-2019-visual-relationship/data\">Relationships Dataset</a>, X is the ROI just big enough to hold two classes, y is the relationship.</p>\n\n<p>Predicting:\n1. Detect objects in the query images, of course, using the detection network.\n1. Get the ROIs(sub-images) which is just big enough to hold two classes. \n1. Classfy the ROIs using the classification network.</p>\n\n<p>I'm not sure if this idea will work, I hope I can spare the time to practice this idea.</p>\n\n<p>One more interesting question:  Are  <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/data\">Detection Dataset</a> and Segmentation Dataset External Data, considering they contain external annotations.</p>",
  "messages": [
    {
      "id": 546121,
      "postDate": "2019-06-06T09:18:17.563Z",
      "content": "<p>I plan to do it this way, if I have time in the next 3 months.</p>\n\n<p>Training:\n- Train a detection network with <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/data\">Detection Dataset</a>. After this, we can get the bbox easily.\n- Train a classification network with the <a href=\"https://www.kaggle.com/c/open-images-2019-visual-relationship/data\">Relationships Dataset</a>, X is the ROI just big enough to hold two classes, y is the relationship.</p>\n\n<p>Predicting:\n1. Detect objects in the query images, of course, using the detection network.\n1. Get the ROIs(sub-images) which is just big enough to hold two classes. \n1. Classfy the ROIs using the classification network.</p>\n\n<p>I'm not sure if this idea will work, I hope I can spare the time to practice this idea.</p>\n\n<p>One more interesting question:  Are  <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/data\">Detection Dataset</a> and Segmentation Dataset External Data, considering they contain external annotations.</p>",
      "rawMarkdown": "I plan to do it this way, if I have time in the next 3 months.\n\nTraining:\n- Train a detection network with [Detection Dataset](https://www.kaggle.com/c/open-images-2019-object-detection/data). After this, we can get the bbox easily.\n- Train a classification network with the [Relationships Dataset](https://www.kaggle.com/c/open-images-2019-visual-relationship/data), X is the ROI just big enough to hold two classes, y is the relationship.\n\nPredicting:\n1. Detect objects in the query images, of course, using the detection network.\n1. Get the ROIs(sub-images) which is just big enough to hold two classes. \n1. Classfy the ROIs using the classification network.\n\nI'm not sure if this idea will work, I hope I can spare the time to practice this idea.\n\nOne more interesting question:  Are  [Detection Dataset](https://www.kaggle.com/c/open-images-2019-object-detection/data) and Segmentation Dataset External Data, considering they contain external annotations.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "546121": "I plan to do it this way, if I have time in the next 3 months.\n\nTraining:\n- Train a detection network with [Detection Dataset](https://www.kaggle.com/c/open-images-2019-object-detection/data). After this, we can get the bbox easily.\n- Train a classification network with the [Relationships Dataset](https://www.kaggle.com/c/open-images-2019-visual-relationship/data), X is the ROI just big enough to hold two classes, y is the relationship.\n\nPredicting:\n1. Detect objects in the query images, of course, using the detection network.\n1. Get the ROIs(sub-images) which is just big enough to hold two classes. \n1. Classfy the ROIs using the classification network.\n\nI'm not sure if this idea will work, I hope I can spare the time to practice this idea.\n\nOne more interesting question:  Are  [Detection Dataset](https://www.kaggle.com/c/open-images-2019-object-detection/data) and Segmentation Dataset External Data, considering they contain external annotations."
  }
}