{
  "id": 112204,
  "title": "Effect of image size on classification",
  "url": "/competitions/kuzushiji-recognition/discussion/112204",
  "author_name": "",
  "post_date": "2019-10-11T12:22:41.507611400Z",
  "votes": 6,
  "comment_count": 4,
  "views": 0,
  "content": "<p>In our experimentation we found that going from 32X32, 64X64, 128X128 and to 224X224 sizes accuracy both increases in CV and LB. Any suggestions on the optimum image size to train the classifier??</p>",
  "messages": [
    {
      "id": "646556",
      "postDate": "10/11/2019 12:22:41",
      "content": "<p>In our experimentation we found that going from 32X32, 64X64, 128X128 and to 224X224 sizes accuracy both increases in CV and LB. Any suggestions on the optimum image size to train the classifier??</p>",
      "rawMarkdown": "In our experimentation we found that going from 32X32, 64X64, 128X128 and to 224X224 sizes accuracy both increases in CV and LB. Any suggestions on the optimum image size to train the classifier??",
      "votes": null
    },
    {
      "id": "647264",
      "postDate": "10/12/2019 10:03:49",
      "content": "<p>I am currently using 96x96. Also, I have tried 128x128 but haven't noticed any improvement.</p>",
      "rawMarkdown": "I am currently using 96x96. Also, I have tried 128x128 but haven't noticed any improvement.",
      "votes": null
    },
    {
      "id": "647304",
      "postDate": "10/12/2019 12:09:08",
      "content": "<p>I'm not using crops around single characters, instead I train on larger crops from the page, but sizes of characters are mostly in 50-100 px range for me, and increasing them didn't help. Example crop used during training:</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19390%2Fcb82c3d969198c80d3bfd93ba1fbe8df%2FScreenshot%202019-10-12%20at%2015.05.29.jpg?generation=1570882115369984&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "I'm not using crops around single characters, instead I train on larger crops from the page, but sizes of characters are mostly in 50-100 px range for me, and increasing them didn't help. Example crop used during training:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19390%2Fcb82c3d969198c80d3bfd93ba1fbe8df%2FScreenshot%202019-10-12%20at%2015.05.29.jpg?generation=1570882115369984&amp;alt=media)",
      "votes": null
    },
    {
      "id": "648015",
      "postDate": "10/13/2019 16:06:23",
      "content": "<p>How can you train your classifier on this? What would the example label for this be?</p>",
      "rawMarkdown": "How can you train your classifier on this? What would the example label for this be?",
      "votes": null
    },
    {
      "id": "648021",
      "postDate": "10/13/2019 16:23:56",
      "content": "<p>The target is similar to target of detection models, but bounding boxes are fixed, so only class is predicted. Architecture is similar to detection, with roi pooling for each character.</p>",
      "rawMarkdown": "The target is similar to target of detection models, but bounding boxes are fixed, so only class is predicted. Architecture is similar to detection, with roi pooling for each character.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 647264,
      "author_name": "opanichev",
      "author_url": "",
      "post_date": "10/12/2019 10:03:49",
      "content": "<p>I am currently using 96x96. Also, I have tried 128x128 but haven't noticed any improvement.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 647304,
      "author_name": "lopuhin",
      "author_url": "",
      "post_date": "10/12/2019 12:09:08",
      "content": "<p>I'm not using crops around single characters, instead I train on larger crops from the page, but sizes of characters are mostly in 50-100 px range for me, and increasing them didn't help. Example crop used during training:</p>\n\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19390%2Fcb82c3d969198c80d3bfd93ba1fbe8df%2FScreenshot%202019-10-12%20at%2015.05.29.jpg?generation=1570882115369984&amp;alt=media\" alt=\"\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 648015,
          "author_name": "usmannkhan",
          "author_url": "",
          "post_date": "10/13/2019 16:06:23",
          "content": "<p>How can you train your classifier on this? What would the example label for this be?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 648021,
          "author_name": "lopuhin",
          "author_url": "",
          "post_date": "10/13/2019 16:23:56",
          "content": "<p>The target is similar to target of detection models, but bounding boxes are fixed, so only class is predicted. Architecture is similar to detection, with roi pooling for each character.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "646556": "In our experimentation we found that going from 32X32, 64X64, 128X128 and to 224X224 sizes accuracy both increases in CV and LB. Any suggestions on the optimum image size to train the classifier??",
    "647264": "I am currently using 96x96. Also, I have tried 128x128 but haven't noticed any improvement.",
    "647304": "I'm not using crops around single characters, instead I train on larger crops from the page, but sizes of characters are mostly in 50-100 px range for me, and increasing them didn't help. Example crop used during training:\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19390%2Fcb82c3d969198c80d3bfd93ba1fbe8df%2FScreenshot%202019-10-12%20at%2015.05.29.jpg?generation=1570882115369984&amp;alt=media)",
    "648015": "How can you train your classifier on this? What would the example label for this be?",
    "648021": "The target is similar to target of detection models, but bounding boxes are fixed, so only class is predicted. Architecture is similar to detection, with roi pooling for each character."
  },
  "source": "meta"
}