{
  "id": 49946,
  "title": "Extremely low score",
  "url": "/competitions/data-science-bowl-2018/discussion/49946",
  "author_name": "",
  "post_date": "2018-02-17T00:05:11.150735800Z",
  "votes": null,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I used thresholding and classical computer vision techniques to get a score of 0.027. Does anyone know of any reasons that could cause such a low score ? My model is similar to <a href=\"https://www.kaggle.com/stkbailey/teaching-notebook-for-total-imaging-newbies\">this</a></p>",
  "messages": [
    {
      "id": "284363",
      "postDate": "02/17/2018 00:05:11",
      "content": "<p>I used thresholding and classical computer vision techniques to get a score of 0.027. Does anyone know of any reasons that could cause such a low score ? My model is similar to <a href=\"https://www.kaggle.com/stkbailey/teaching-notebook-for-total-imaging-newbies\">this</a></p>",
      "rawMarkdown": "I used thresholding and classical computer vision techniques to get a score of 0.027. Does anyone know of any reasons that could cause such a low score ? My model is similar to [this][1]\n\n\n  [1]: https://www.kaggle.com/stkbailey/teaching-notebook-for-total-imaging-newbies",
      "votes": null
    },
    {
      "id": "285224",
      "postDate": "02/19/2018 16:27:18",
      "content": "<p>Hey bkKaggle,</p>\n\n<p>I don't know the details of the competition (how bad is 0.027?), but I want to try to answer anyway</p>\n\n<p>Did you set apart some of your training data to validate your model?</p>\n\n<ul>\n<li><p><em>No, I didn't validate</em> -&gt; Maybe it was a neural network that failed to converge?</p></li>\n<li><p><em>Yes, I validated and got nice results on validation</em> -&gt; Maybe you mismatched the indexes and predictions when creating your submission file?</p></li>\n</ul>",
      "rawMarkdown": "Hey bkKaggle,\n\nI don't know the details of the competition (how bad is 0.027?), but I want to try to answer anyway\n\nDid you set apart some of your training data to validate your model?\n\n- *No, I didn't validate* -&gt; Maybe it was a neural network that failed to converge?\n\n- *Yes, I validated and got nice results on validation* -&gt; Maybe you mismatched the indexes and predictions when creating your submission file?",
      "votes": null
    },
    {
      "id": "285239",
      "postDate": "02/19/2018 17:06:14",
      "content": "<p>The kernel you linked to doesn't use any deep learning and is more aimed at getting familiar with image processing and the data than getting a high score on the leader board. To score high in this competition you would probably need to create an instance segmentation model like Mask RCNN or something similar.</p>\n\n<p>Allen Goodman who is currently at the top of the leaderboard mentioned that he is using a Mask RCNN model. There is a thread on Mask RCNN here <a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692</a></p>",
      "rawMarkdown": "The kernel you linked to doesn't use any deep learning and is more aimed at getting familiar with image processing and the data than getting a high score on the leader board. To score high in this competition you would probably need to create an instance segmentation model like Mask RCNN or something similar.\n\nAllen Goodman who is currently at the top of the leaderboard mentioned that he is using a Mask RCNN model. There is a thread on Mask RCNN here https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692",
      "votes": null
    },
    {
      "id": "285479",
      "postDate": "02/20/2018 11:33:25",
      "content": "<p>Hey @bkKaggle -- one thing that you might check is whether you are getting the right \"half\" of the thresholded image for your masks. That is, when you threshold your image, you might actually want the bottom half! This line of code in the linked kernel checks that the \"half\" that you want has less data than the background half - and if it does not, then it inverts the mask.</p>\n\n<pre><code># Make sure the larger portion of the mask is considered background\nif np.sum(mask==0) &lt; np.sum(mask==1):\n    mask = np.where(mask, 0, 1)\n</code></pre>",
      "rawMarkdown": "Hey @bkKaggle -- one thing that you might check is whether you are getting the right \"half\" of the thresholded image for your masks. That is, when you threshold your image, you might actually want the bottom half! This line of code in the linked kernel checks that the \"half\" that you want has less data than the background half - and if it does not, then it inverts the mask.\n\n\n    # Make sure the larger portion of the mask is considered background\n    if np.sum(mask==0) &lt; np.sum(mask==1):\n        mask = np.where(mask, 0, 1)",
      "votes": null
    },
    {
      "id": "285681",
      "postDate": "02/20/2018 20:38:33",
      "content": "<p>Did you look at the results you get with the test set? The methods in the notebook you link to will give at best 0.3 precision for the easily segmentable images, and probably 0 for hard images.</p>",
      "rawMarkdown": "Did you look at the results you get with the test set? The methods in the notebook you link to will give at best 0.3 precision for the easily segmentable images, and probably 0 for hard images.",
      "votes": null
    },
    {
      "id": "288212",
      "postDate": "02/27/2018 00:30:13",
      "content": "<p>This solved the problem. Thank you</p>",
      "rawMarkdown": "This solved the problem. Thank you",
      "votes": null
    },
    {
      "id": "291504",
      "postDate": "03/06/2018 08:50:58",
      "content": "<p>looks like i have same problem here</p>",
      "rawMarkdown": "looks like i have same problem here",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 285224,
      "author_name": "jellevde",
      "author_url": "",
      "post_date": "02/19/2018 16:27:18",
      "content": "<p>Hey bkKaggle,</p>\n\n<p>I don't know the details of the competition (how bad is 0.027?), but I want to try to answer anyway</p>\n\n<p>Did you set apart some of your training data to validate your model?</p>\n\n<ul>\n<li><p><em>No, I didn't validate</em> -&gt; Maybe it was a neural network that failed to converge?</p></li>\n<li><p><em>Yes, I validated and got nice results on validation</em> -&gt; Maybe you mismatched the indexes and predictions when creating your submission file?</p></li>\n</ul>",
      "votes": null,
      "replies": []
    },
    {
      "id": 285239,
      "author_name": "jackvial",
      "author_url": "",
      "post_date": "02/19/2018 17:06:14",
      "content": "<p>The kernel you linked to doesn't use any deep learning and is more aimed at getting familiar with image processing and the data than getting a high score on the leader board. To score high in this competition you would probably need to create an instance segmentation model like Mask RCNN or something similar.</p>\n\n<p>Allen Goodman who is currently at the top of the leaderboard mentioned that he is using a Mask RCNN model. There is a thread on Mask RCNN here <a href=\"https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692\">https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 285479,
      "author_name": "stkbailey",
      "author_url": "",
      "post_date": "02/20/2018 11:33:25",
      "content": "<p>Hey @bkKaggle -- one thing that you might check is whether you are getting the right \"half\" of the thresholded image for your masks. That is, when you threshold your image, you might actually want the bottom half! This line of code in the linked kernel checks that the \"half\" that you want has less data than the background half - and if it does not, then it inverts the mask.</p>\n\n<pre><code># Make sure the larger portion of the mask is considered background\nif np.sum(mask==0) &lt; np.sum(mask==1):\n    mask = np.where(mask, 0, 1)\n</code></pre>",
      "votes": null,
      "replies": [
        {
          "id": 288212,
          "author_name": "bkkaggle",
          "author_url": "",
          "post_date": "02/27/2018 00:30:13",
          "content": "<p>This solved the problem. Thank you</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 285681,
      "author_name": "tichyijon",
      "author_url": "",
      "post_date": "02/20/2018 20:38:33",
      "content": "<p>Did you look at the results you get with the test set? The methods in the notebook you link to will give at best 0.3 precision for the easily segmentable images, and probably 0 for hard images.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 291504,
      "author_name": "sayaroy",
      "author_url": "",
      "post_date": "03/06/2018 08:50:58",
      "content": "<p>looks like i have same problem here</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "284363": "I used thresholding and classical computer vision techniques to get a score of 0.027. Does anyone know of any reasons that could cause such a low score ? My model is similar to [this][1]\n\n\n  [1]: https://www.kaggle.com/stkbailey/teaching-notebook-for-total-imaging-newbies",
    "285224": "Hey bkKaggle,\n\nI don't know the details of the competition (how bad is 0.027?), but I want to try to answer anyway\n\nDid you set apart some of your training data to validate your model?\n\n- *No, I didn't validate* -&gt; Maybe it was a neural network that failed to converge?\n\n- *Yes, I validated and got nice results on validation* -&gt; Maybe you mismatched the indexes and predictions when creating your submission file?",
    "285239": "The kernel you linked to doesn't use any deep learning and is more aimed at getting familiar with image processing and the data than getting a high score on the leader board. To score high in this competition you would probably need to create an instance segmentation model like Mask RCNN or something similar.\n\nAllen Goodman who is currently at the top of the leaderboard mentioned that he is using a Mask RCNN model. There is a thread on Mask RCNN here https://www.kaggle.com/c/data-science-bowl-2018/discussion/49692",
    "285479": "Hey @bkKaggle -- one thing that you might check is whether you are getting the right \"half\" of the thresholded image for your masks. That is, when you threshold your image, you might actually want the bottom half! This line of code in the linked kernel checks that the \"half\" that you want has less data than the background half - and if it does not, then it inverts the mask.\n\n\n    # Make sure the larger portion of the mask is considered background\n    if np.sum(mask==0) &lt; np.sum(mask==1):\n        mask = np.where(mask, 0, 1)",
    "285681": "Did you look at the results you get with the test set? The methods in the notebook you link to will give at best 0.3 precision for the easily segmentable images, and probably 0 for hard images.",
    "288212": "This solved the problem. Thank you",
    "291504": "looks like i have same problem here"
  },
  "source": "meta"
}