{
  "id": 69412,
  "title": "Why would you set a maximum number of ground truth instances?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/69412",
  "author_name": "",
  "post_date": "2018-10-23T16:58:51.837765900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I'm reading about the masked RCNN model and trying to learn what all of the config options do. One of the options is:</p>\n\n<p>MAX_GT_INSTANCES = 3</p>\n\n<p>followed by :\nDETECTION_MAX_INSTANCES = 3</p>\n\n<p>If I'm reading this correctly, doesn't this say \"Take at most three supplied bounding boxes (per image) to train with and when predicting give at most 3 predicted areas. \"</p>\n\n<p>Wouldn't you want to use all of the GT bonding boxes possible? (or what would happen if you made this a large number?)  </p>\n\n<p>As for the max detected instances, this makes more sense to prevent false positives, but wouldn't it be better to use the detection confidence to weed out false positives? </p>\n\n<p>I ask this as a total novice to the field. </p>\n\n<p>I've been studying the code from: <a href=\"https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155\">https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155</a></p>\n\n<p>Thanks.</p>",
  "messages": [
    {
      "id": "408967",
      "postDate": "10/23/2018 16:58:51",
      "content": "<p>I'm reading about the masked RCNN model and trying to learn what all of the config options do. One of the options is:</p>\n\n<p>MAX_GT_INSTANCES = 3</p>\n\n<p>followed by :\nDETECTION_MAX_INSTANCES = 3</p>\n\n<p>If I'm reading this correctly, doesn't this say \"Take at most three supplied bounding boxes (per image) to train with and when predicting give at most 3 predicted areas. \"</p>\n\n<p>Wouldn't you want to use all of the GT bonding boxes possible? (or what would happen if you made this a large number?)  </p>\n\n<p>As for the max detected instances, this makes more sense to prevent false positives, but wouldn't it be better to use the detection confidence to weed out false positives? </p>\n\n<p>I ask this as a total novice to the field. </p>\n\n<p>I've been studying the code from: <a href=\"https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155\">https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155</a></p>\n\n<p>Thanks.</p>",
      "rawMarkdown": "I'm reading about the masked RCNN model and trying to learn what all of the config options do. One of the options is:\n\nMAX_GT_INSTANCES = 3\n\nfollowed by :\nDETECTION_MAX_INSTANCES = 3\n\nIf I'm reading this correctly, doesn't this say \"Take at most three supplied bounding boxes (per image) to train with and when predicting give at most 3 predicted areas. \"\n\nWouldn't you want to use all of the GT bonding boxes possible? (or what would happen if you made this a large number?)  \n\nAs for the max detected instances, this makes more sense to prevent false positives, but wouldn't it be better to use the detection confidence to weed out false positives? \n\nI ask this as a total novice to the field. \n\n\nI've been studying the code from: https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155\n\n\nThanks.",
      "votes": null
    },
    {
      "id": "409279",
      "postDate": "10/24/2018 03:06:59",
      "content": "<p>you got it.\nyou can modify the code(basically in model.py) with your idea and tell us the answer</p>",
      "rawMarkdown": "you got it.\nyou can modify the code(basically in model.py) with your idea and tell us the answer",
      "votes": null
    },
    {
      "id": "409425",
      "postDate": "10/24/2018 09:16:52",
      "content": "<p>the current version uses MAX_GT_INSTANCES = 4 as there aren't any images with more than 4 boxes anyway\nDETECTION_MAX_INSTANCES = 2 might actually improve your LB as the model is not perfect and most images have max 2 boxes...</p>",
      "rawMarkdown": "the current version uses MAX_GT_INSTANCES = 4 as there aren't any images with more than 4 boxes anyway\nDETECTION_MAX_INSTANCES = 2 might actually improve your LB as the model is not perfect and most images have max 2 boxes...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 409279,
      "author_name": "atom1231",
      "author_url": "",
      "post_date": "10/24/2018 03:06:59",
      "content": "<p>you got it.\nyou can modify the code(basically in model.py) with your idea and tell us the answer</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 409425,
      "author_name": "hmendonca",
      "author_url": "",
      "post_date": "10/24/2018 09:16:52",
      "content": "<p>the current version uses MAX_GT_INSTANCES = 4 as there aren't any images with more than 4 boxes anyway\nDETECTION_MAX_INSTANCES = 2 might actually improve your LB as the model is not perfect and most images have max 2 boxes...</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "408967": "I'm reading about the masked RCNN model and trying to learn what all of the config options do. One of the options is:\n\nMAX_GT_INSTANCES = 3\n\nfollowed by :\nDETECTION_MAX_INSTANCES = 3\n\nIf I'm reading this correctly, doesn't this say \"Take at most three supplied bounding boxes (per image) to train with and when predicting give at most 3 predicted areas. \"\n\nWouldn't you want to use all of the GT bonding boxes possible? (or what would happen if you made this a large number?)  \n\nAs for the max detected instances, this makes more sense to prevent false positives, but wouldn't it be better to use the detection confidence to weed out false positives? \n\nI ask this as a total novice to the field. \n\n\nI've been studying the code from: https://www.kaggle.com/hmendonca/mask-rcnn-and-coco-transfer-learning-lb-0-155\n\n\nThanks.",
    "409279": "you got it.\nyou can modify the code(basically in model.py) with your idea and tell us the answer",
    "409425": "the current version uses MAX_GT_INSTANCES = 4 as there aren't any images with more than 4 boxes anyway\nDETECTION_MAX_INSTANCES = 2 might actually improve your LB as the model is not perfect and most images have max 2 boxes..."
  },
  "source": "meta"
}