{
  "id": 88089,
  "title": "Dataset Image Quality",
  "url": "/competitions/histopathologic-cancer-detection/discussion/88089",
  "author_name": "",
  "post_date": "2019-04-05T15:43:12.526660500Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p><img src=\"https://imgur.com/a/8ph2WLN\" alt=\"32x32 area Compare Image quality\"> \n... unfortunately I can't seem to get this image to show up on kaggle so you need to open it directly</p>\n\n<p>Some (most?) who have worked with the Patch Camelyon dataset on kaggle have been frustrated by the image quality, both the resolution and the nasty jpeg artifacts.  Since the dataset was derived from Camelyon16 I decided to look into this further. In the attached image the left column is the 32x32 center window of samples of the source data from the competition. The second column is that portion of the image rescaled to 224x224 as used by some participants in the competition.  The third column is the source data from the original dataset and the forth is that data rescaled again to 224x224.</p>\n\n<p>Observations: Column1 is very poor quality. Also when separated as HSL it is very easy to see that the data contains ugly jpeg artifacts.\nColumn2, (a non-integer scaling of Column1) shows improvement to the shapes of features. Some of the better performing models used rescaling, here we can see how rescaling has helped.\nColumn3 shows what is actually available in the original data, Which is 4x the quality of that released for the kaggle competition. \nColumn4 shows again a non-integer scaling of this data which also seems to helps to improve image quality.</p>\n\n<p>for reference these images are derived from:\n<code>\n62ae3a6e3ba8355f617455473bf4d8eec423a834.tif\n5d3753be0ddbe2ff18e5b2b07ecb4618540857a3.tif\n3b110a231103153bc461939c4eb7ff89bd9bdcdb.tif\n6414bedd74bceb877cfd813b4ff6219908307839.tif\nf91e889c962fb8dee3dd6293e1d68340682ff9c5.tif\n59292b9df82ef0e64552ced2bc89bbac5d00e92c.tif\n</code>\nIt would be nice to have a re-release of the dataset with better quality so we could build improved models. I can build this dataset myself but my internet service makes image transfer slow.</p>",
  "messages": [
    {
      "id": "508098",
      "postDate": "04/05/2019 15:43:12",
      "content": "<p><img src=\"https://imgur.com/a/8ph2WLN\" alt=\"32x32 area Compare Image quality\"> \n... unfortunately I can't seem to get this image to show up on kaggle so you need to open it directly</p>\n\n<p>Some (most?) who have worked with the Patch Camelyon dataset on kaggle have been frustrated by the image quality, both the resolution and the nasty jpeg artifacts.  Since the dataset was derived from Camelyon16 I decided to look into this further. In the attached image the left column is the 32x32 center window of samples of the source data from the competition. The second column is that portion of the image rescaled to 224x224 as used by some participants in the competition.  The third column is the source data from the original dataset and the forth is that data rescaled again to 224x224.</p>\n\n<p>Observations: Column1 is very poor quality. Also when separated as HSL it is very easy to see that the data contains ugly jpeg artifacts.\nColumn2, (a non-integer scaling of Column1) shows improvement to the shapes of features. Some of the better performing models used rescaling, here we can see how rescaling has helped.\nColumn3 shows what is actually available in the original data, Which is 4x the quality of that released for the kaggle competition. \nColumn4 shows again a non-integer scaling of this data which also seems to helps to improve image quality.</p>\n\n<p>for reference these images are derived from:\n<code>\n62ae3a6e3ba8355f617455473bf4d8eec423a834.tif\n5d3753be0ddbe2ff18e5b2b07ecb4618540857a3.tif\n3b110a231103153bc461939c4eb7ff89bd9bdcdb.tif\n6414bedd74bceb877cfd813b4ff6219908307839.tif\nf91e889c962fb8dee3dd6293e1d68340682ff9c5.tif\n59292b9df82ef0e64552ced2bc89bbac5d00e92c.tif\n</code>\nIt would be nice to have a re-release of the dataset with better quality so we could build improved models. I can build this dataset myself but my internet service makes image transfer slow.</p>",
      "rawMarkdown": "![32x32 area Compare Image quality](https://imgur.com/a/8ph2WLN) \n... unfortunately I can't seem to get this image to show up on kaggle so you need to open it directly\n\nSome (most?) who have worked with the Patch Camelyon dataset on kaggle have been frustrated by the image quality, both the resolution and the nasty jpeg artifacts.  Since the dataset was derived from Camelyon16 I decided to look into this further. In the attached image the left column is the 32x32 center window of samples of the source data from the competition. The second column is that portion of the image rescaled to 224x224 as used by some participants in the competition.  The third column is the source data from the original dataset and the forth is that data rescaled again to 224x224.\n\nObservations: Column1 is very poor quality. Also when separated as HSL it is very easy to see that the data contains ugly jpeg artifacts.\nColumn2, (a non-integer scaling of Column1) shows improvement to the shapes of features. Some of the better performing models used rescaling, here we can see how rescaling has helped.\nColumn3 shows what is actually available in the original data, Which is 4x the quality of that released for the kaggle competition. \nColumn4 shows again a non-integer scaling of this data which also seems to helps to improve image quality.\n\nfor reference these images are derived from:\n```\n62ae3a6e3ba8355f617455473bf4d8eec423a834.tif\n5d3753be0ddbe2ff18e5b2b07ecb4618540857a3.tif\n3b110a231103153bc461939c4eb7ff89bd9bdcdb.tif\n6414bedd74bceb877cfd813b4ff6219908307839.tif\nf91e889c962fb8dee3dd6293e1d68340682ff9c5.tif\n59292b9df82ef0e64552ced2bc89bbac5d00e92c.tif\n```\nIt would be nice to have a re-release of the dataset with better quality so we could build improved models. I can build this dataset myself but my internet service makes image transfer slow.",
      "votes": null
    },
    {
      "id": "508221",
      "postDate": "04/05/2019 20:03:03",
      "content": "<p>Oh yeah. Read my post about it too. Thanks for sharing. Anyway, fighting with bad weapons improve your skills! </p>",
      "rawMarkdown": "Oh yeah. Read my post about it too. Thanks for sharing. Anyway, fighting with bad weapons improve your skills!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 508221,
      "author_name": "simoninparis",
      "author_url": "",
      "post_date": "04/05/2019 20:03:03",
      "content": "<p>Oh yeah. Read my post about it too. Thanks for sharing. Anyway, fighting with bad weapons improve your skills! </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "508098": "![32x32 area Compare Image quality](https://imgur.com/a/8ph2WLN) \n... unfortunately I can't seem to get this image to show up on kaggle so you need to open it directly\n\nSome (most?) who have worked with the Patch Camelyon dataset on kaggle have been frustrated by the image quality, both the resolution and the nasty jpeg artifacts.  Since the dataset was derived from Camelyon16 I decided to look into this further. In the attached image the left column is the 32x32 center window of samples of the source data from the competition. The second column is that portion of the image rescaled to 224x224 as used by some participants in the competition.  The third column is the source data from the original dataset and the forth is that data rescaled again to 224x224.\n\nObservations: Column1 is very poor quality. Also when separated as HSL it is very easy to see that the data contains ugly jpeg artifacts.\nColumn2, (a non-integer scaling of Column1) shows improvement to the shapes of features. Some of the better performing models used rescaling, here we can see how rescaling has helped.\nColumn3 shows what is actually available in the original data, Which is 4x the quality of that released for the kaggle competition. \nColumn4 shows again a non-integer scaling of this data which also seems to helps to improve image quality.\n\nfor reference these images are derived from:\n```\n62ae3a6e3ba8355f617455473bf4d8eec423a834.tif\n5d3753be0ddbe2ff18e5b2b07ecb4618540857a3.tif\n3b110a231103153bc461939c4eb7ff89bd9bdcdb.tif\n6414bedd74bceb877cfd813b4ff6219908307839.tif\nf91e889c962fb8dee3dd6293e1d68340682ff9c5.tif\n59292b9df82ef0e64552ced2bc89bbac5d00e92c.tif\n```\nIt would be nice to have a re-release of the dataset with better quality so we could build improved models. I can build this dataset myself but my internet service makes image transfer slow.",
    "508221": "Oh yeah. Read my post about it too. Thanks for sharing. Anyway, fighting with bad weapons improve your skills!"
  },
  "source": "meta"
}