{
  "id": 89637,
  "title": "Can Random Erase method improve results",
  "url": "/competitions/imet-2019-fgvc6/discussion/89637",
  "author_name": "",
  "post_date": "2019-04-16T11:51:51.667418300Z",
  "votes": 12,
  "comment_count": 17,
  "views": 0,
  "content": "<p>In order to increase the robustness of the model, We often do some data ugmentation like this:\n1.Random crop \n2.RandomMirror \n3.SubtractMeans \n4.Reszie ...\nBut here , I wonder if I can add this method (Random Erase) when I think about it to increase the robustness of the model.\nThe effect is as shown this: (<a href=\"https://www.kaggle.com/suyukun/random-erase-pic\">https://www.kaggle.com/suyukun/random-erase-pic</a>)</p>",
  "messages": [
    {
      "id": "517698",
      "postDate": "04/16/2019 11:51:51",
      "content": "<p>In order to increase the robustness of the model, We often do some data ugmentation like this:\n1.Random crop \n2.RandomMirror \n3.SubtractMeans \n4.Reszie ...\nBut here , I wonder if I can add this method (Random Erase) when I think about it to increase the robustness of the model.\nThe effect is as shown this: (<a href=\"https://www.kaggle.com/suyukun/random-erase-pic\">https://www.kaggle.com/suyukun/random-erase-pic</a>)</p>",
      "rawMarkdown": "In order to increase the robustness of the model, We often do some data ugmentation like this:\n1.Random crop \n2.RandomMirror \n3.SubtractMeans \n4.Reszie ...\nBut here , I wonder if I can add this method (Random Erase) when I think about it to increase the robustness of the model.\nThe effect is as shown this: (https://www.kaggle.com/suyukun/random-erase-pic)",
      "votes": null
    },
    {
      "id": "517768",
      "postDate": "04/16/2019 13:55:52",
      "content": "<p>Sounds nice </p>",
      "rawMarkdown": "Sounds nice",
      "votes": null
    },
    {
      "id": "517788",
      "postDate": "04/16/2019 14:34:26",
      "content": "<p>thx</p>",
      "rawMarkdown": "thx",
      "votes": null
    },
    {
      "id": "517810",
      "postDate": "04/16/2019 15:09:10",
      "content": "<p>Thanks for sharing technique! I found a related paper. Good.\nRandom Erasing Data Augmentation\n<a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a></p>",
      "rawMarkdown": "Thanks for sharing technique! I found a related paper. Good.\nRandom Erasing Data Augmentation\nhttps://arxiv.org/pdf/1708.04896.pdf",
      "votes": null
    },
    {
      "id": "517821",
      "postDate": "04/16/2019 15:22:05",
      "content": "<p>You are welcome！I am adding this  method to verify the feasibility of the model. And this can be used not only in image classification but also objection detection</p>",
      "rawMarkdown": "You are welcome！I am adding this  method to verify the feasibility of the model. And this can be used not only in image classification but also objection detection",
      "votes": null
    },
    {
      "id": "517909",
      "postDate": "04/16/2019 16:56:35",
      "content": "<p>How knows? Lets have a try</p>",
      "rawMarkdown": "How knows? Lets have a try",
      "votes": null
    },
    {
      "id": "517982",
      "postDate": "04/16/2019 18:19:34",
      "content": "<p>Try it! :)</p>",
      "rawMarkdown": "Try it! :)",
      "votes": null
    },
    {
      "id": "518211",
      "postDate": "04/17/2019 01:20:59",
      "content": "<p>According to this paper : <a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a>, it's idea can achieve the robustness of the model. Let's try</p>",
      "rawMarkdown": "According to this paper : https://arxiv.org/pdf/1708.04896.pdf, it's idea can achieve the robustness of the model. Let's try",
      "votes": null
    },
    {
      "id": "518212",
      "postDate": "04/17/2019 01:21:35",
      "content": "<p>yep! Thank you for your appreciation.</p>",
      "rawMarkdown": "yep! Thank you for your appreciation.",
      "votes": null
    },
    {
      "id": "519255",
      "postDate": "04/18/2019 16:13:21",
      "content": "<p>This method sounds pretty good, a bit like a cover</p>",
      "rawMarkdown": "This method sounds pretty good, a bit like a cover",
      "votes": null
    },
    {
      "id": "519435",
      "postDate": "04/19/2019 00:30:29",
      "content": "<p>maybe it worths a try🧐</p>",
      "rawMarkdown": "maybe it worths a try🧐",
      "votes": null
    },
    {
      "id": "520609",
      "postDate": "04/21/2019 12:09:06",
      "content": "<p>Does dropout layer have the same effect? </p>",
      "rawMarkdown": "Does dropout layer have the same effect?",
      "votes": null
    },
    {
      "id": "520705",
      "postDate": "04/21/2019 15:56:59",
      "content": "<p>Is this different from cutout?</p>",
      "rawMarkdown": "Is this different from cutout?",
      "votes": null
    },
    {
      "id": "520925",
      "postDate": "04/22/2019 01:59:34",
      "content": "<p>i have not try yet</p>",
      "rawMarkdown": "i have not try yet",
      "votes": null
    },
    {
      "id": "520927",
      "postDate": "04/22/2019 02:08:43",
      "content": "<p>Referring to these two papers, I think their ideas and motivations are similar, but there are differences in the details of implementation.\n<a href=\"https://arxiv.org/pdf/1708.04552.pdf\">https://arxiv.org/pdf/1708.04552.pdf</a>\n<a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a></p>",
      "rawMarkdown": "Referring to these two papers, I think their ideas and motivations are similar, but there are differences in the details of implementation.\nhttps://arxiv.org/pdf/1708.04552.pdf\nhttps://arxiv.org/pdf/1708.04896.pdf",
      "votes": null
    },
    {
      "id": "520958",
      "postDate": "04/22/2019 04:25:55",
      "content": "<p>Can <a href=\"https://arxiv.org/pdf/1710.09412.pdf\">mixup</a> method improve results?</p>",
      "rawMarkdown": "Can [mixup](https://arxiv.org/pdf/1710.09412.pdf) method improve results?",
      "votes": null
    },
    {
      "id": "521097",
      "postDate": "04/22/2019 10:10:18",
      "content": "<p>Thx! Mixup and Paring Samples can be consider good choices</p>",
      "rawMarkdown": "Thx! Mixup and Paring Samples can be consider good choices",
      "votes": null
    },
    {
      "id": "521576",
      "postDate": "04/23/2019 04:04:55",
      "content": "<p>At first, I had no idea if it gonna help or not. However, after some prediction analysis I could see one of the frequent mistakes of my model is that *<em>it usually detect 'a man with cover' (e.g. Greece or Renaissance  clothes) as 'tag:women' *</em>. For example,</p>\n\n<p><img src=\"https://i.ibb.co/TByr4Pk/1555992451162.jpg\" alt=\"\"></p>\n\n<p>Note that there are lots of pictures like this, and tag:men and tag:women are the most common.</p>\n\n<p>Therefore, it may be possible that by somehow erasing this cover, the model will get more understanding of what is actually 'women' </p>\n\n<p>PS. And yes, this may be similar to 'CoarseDropout' by imgaug, see examples at\n<a href=\"https://github.com/aleju/imgaug\">https://github.com/aleju/imgaug</a></p>",
      "rawMarkdown": "At first, I had no idea if it gonna help or not. However, after some prediction analysis I could see one of the frequent mistakes of my model is that **it usually detect 'a man with cover' (e.g. Greece or Renaissance  clothes) as 'tag:women' **. For example,\n\n![](https://i.ibb.co/TByr4Pk/1555992451162.jpg)\n\nNote that there are lots of pictures like this, and tag:men and tag:women are the most common.\n\nTherefore, it may be possible that by somehow erasing this cover, the model will get more understanding of what is actually 'women' \n\nPS. And yes, this may be similar to 'CoarseDropout' by imgaug, see examples at\nhttps://github.com/aleju/imgaug",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 517768,
      "author_name": "fudashuai",
      "author_url": "",
      "post_date": "04/16/2019 13:55:52",
      "content": "<p>Sounds nice </p>",
      "votes": null,
      "replies": [
        {
          "id": 517788,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/16/2019 14:34:26",
          "content": "<p>thx</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 517810,
      "author_name": "dhaqui",
      "author_url": "",
      "post_date": "04/16/2019 15:09:10",
      "content": "<p>Thanks for sharing technique! I found a related paper. Good.\nRandom Erasing Data Augmentation\n<a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 517821,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/16/2019 15:22:05",
          "content": "<p>You are welcome！I am adding this  method to verify the feasibility of the model. And this can be used not only in image classification but also objection detection</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 517909,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "04/16/2019 16:56:35",
      "content": "<p>How knows? Lets have a try</p>",
      "votes": null,
      "replies": [
        {
          "id": 518211,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/17/2019 01:20:59",
          "content": "<p>According to this paper : <a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a>, it's idea can achieve the robustness of the model. Let's try</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 517982,
      "author_name": "alexanderliao",
      "author_url": "",
      "post_date": "04/16/2019 18:19:34",
      "content": "<p>Try it! :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 518212,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/17/2019 01:21:35",
          "content": "<p>yep! Thank you for your appreciation.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 519255,
      "author_name": "powercode",
      "author_url": "",
      "post_date": "04/18/2019 16:13:21",
      "content": "<p>This method sounds pretty good, a bit like a cover</p>",
      "votes": null,
      "replies": [
        {
          "id": 519435,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/19/2019 00:30:29",
          "content": "<p>maybe it worths a try🧐</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 520609,
      "author_name": "martinid",
      "author_url": "",
      "post_date": "04/21/2019 12:09:06",
      "content": "<p>Does dropout layer have the same effect? </p>",
      "votes": null,
      "replies": [
        {
          "id": 520925,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/22/2019 01:59:34",
          "content": "<p>i have not try yet</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 520705,
      "author_name": "interneuron",
      "author_url": "",
      "post_date": "04/21/2019 15:56:59",
      "content": "<p>Is this different from cutout?</p>",
      "votes": null,
      "replies": [
        {
          "id": 520927,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/22/2019 02:08:43",
          "content": "<p>Referring to these two papers, I think their ideas and motivations are similar, but there are differences in the details of implementation.\n<a href=\"https://arxiv.org/pdf/1708.04552.pdf\">https://arxiv.org/pdf/1708.04552.pdf</a>\n<a href=\"https://arxiv.org/pdf/1708.04896.pdf\">https://arxiv.org/pdf/1708.04896.pdf</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 520958,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "04/22/2019 04:25:55",
      "content": "<p>Can <a href=\"https://arxiv.org/pdf/1710.09412.pdf\">mixup</a> method improve results?</p>",
      "votes": null,
      "replies": [
        {
          "id": 521097,
          "author_name": "suyukun",
          "author_url": "",
          "post_date": "04/22/2019 10:10:18",
          "content": "<p>Thx! Mixup and Paring Samples can be consider good choices</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 521576,
      "author_name": "ratthachat",
      "author_url": "",
      "post_date": "04/23/2019 04:04:55",
      "content": "<p>At first, I had no idea if it gonna help or not. However, after some prediction analysis I could see one of the frequent mistakes of my model is that *<em>it usually detect 'a man with cover' (e.g. Greece or Renaissance  clothes) as 'tag:women' *</em>. For example,</p>\n\n<p><img src=\"https://i.ibb.co/TByr4Pk/1555992451162.jpg\" alt=\"\"></p>\n\n<p>Note that there are lots of pictures like this, and tag:men and tag:women are the most common.</p>\n\n<p>Therefore, it may be possible that by somehow erasing this cover, the model will get more understanding of what is actually 'women' </p>\n\n<p>PS. And yes, this may be similar to 'CoarseDropout' by imgaug, see examples at\n<a href=\"https://github.com/aleju/imgaug\">https://github.com/aleju/imgaug</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "517698": "In order to increase the robustness of the model, We often do some data ugmentation like this:\n1.Random crop \n2.RandomMirror \n3.SubtractMeans \n4.Reszie ...\nBut here , I wonder if I can add this method (Random Erase) when I think about it to increase the robustness of the model.\nThe effect is as shown this: (https://www.kaggle.com/suyukun/random-erase-pic)",
    "517768": "Sounds nice",
    "517788": "thx",
    "517810": "Thanks for sharing technique! I found a related paper. Good.\nRandom Erasing Data Augmentation\nhttps://arxiv.org/pdf/1708.04896.pdf",
    "517821": "You are welcome！I am adding this  method to verify the feasibility of the model. And this can be used not only in image classification but also objection detection",
    "517909": "How knows? Lets have a try",
    "517982": "Try it! :)",
    "518211": "According to this paper : https://arxiv.org/pdf/1708.04896.pdf, it's idea can achieve the robustness of the model. Let's try",
    "518212": "yep! Thank you for your appreciation.",
    "519255": "This method sounds pretty good, a bit like a cover",
    "519435": "maybe it worths a try🧐",
    "520609": "Does dropout layer have the same effect?",
    "520705": "Is this different from cutout?",
    "520925": "i have not try yet",
    "520927": "Referring to these two papers, I think their ideas and motivations are similar, but there are differences in the details of implementation.\nhttps://arxiv.org/pdf/1708.04552.pdf\nhttps://arxiv.org/pdf/1708.04896.pdf",
    "520958": "Can [mixup](https://arxiv.org/pdf/1710.09412.pdf) method improve results?",
    "521097": "Thx! Mixup and Paring Samples can be consider good choices",
    "521576": "At first, I had no idea if it gonna help or not. However, after some prediction analysis I could see one of the frequent mistakes of my model is that **it usually detect 'a man with cover' (e.g. Greece or Renaissance  clothes) as 'tag:women' **. For example,\n\n![](https://i.ibb.co/TByr4Pk/1555992451162.jpg)\n\nNote that there are lots of pictures like this, and tag:men and tag:women are the most common.\n\nTherefore, it may be possible that by somehow erasing this cover, the model will get more understanding of what is actually 'women' \n\nPS. And yes, this may be similar to 'CoarseDropout' by imgaug, see examples at\nhttps://github.com/aleju/imgaug"
  },
  "source": "meta"
}