{
  "id": 38879,
  "title": "What augmentation methods do you use?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/38879",
  "author_name": "",
  "post_date": "2017-09-01T19:55:46.860217400Z",
  "votes": 8,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I collected a few methods. <a href=\"https://www.kaggle.com/gaborfodor/augmentation-methods\">https://www.kaggle.com/gaborfodor/augmentation-methods</a></p>",
  "messages": [
    {
      "id": "218011",
      "postDate": "09/01/2017 19:55:46",
      "content": "<p>I collected a few methods. <a href=\"https://www.kaggle.com/gaborfodor/augmentation-methods\">https://www.kaggle.com/gaborfodor/augmentation-methods</a></p>",
      "rawMarkdown": "I collected a few methods. https://www.kaggle.com/gaborfodor/augmentation-methods",
      "votes": null
    },
    {
      "id": "218138",
      "postDate": "09/02/2017 13:51:21",
      "content": "<p>I'm just using flip. Have you quantified the effect of other augmentation methods?</p>",
      "rawMarkdown": "I'm just using flip. Have you quantified the effect of other augmentation methods?",
      "votes": null
    },
    {
      "id": "218158",
      "postDate": "09/02/2017 16:44:31",
      "content": "<p>Nothing conclusive yet. I am still playing with different network structures with lower resolution. My current training time is just too long for proper experiments.</p>",
      "rawMarkdown": "Nothing conclusive yet. I am still playing with different network structures with lower resolution. My current training time is just too long for proper experiments.",
      "votes": null
    },
    {
      "id": "218171",
      "postDate": "09/02/2017 18:58:23",
      "content": "<p>Hi, beluga. I also use flip and it looks most effective so far, other augmentation methods makes result much worse. I am going to try some ideas from GANs, I have very strong overfitting. Do you think it could help to cope with overfitting?  </p>",
      "rawMarkdown": "Hi, beluga. I also use flip and it looks most effective so far, other augmentation methods makes result much worse. I am going to try some ideas from GANs, I have very strong overfitting. Do you think it could help to cope with overfitting?",
      "votes": null
    },
    {
      "id": "218342",
      "postDate": "09/03/2017 20:30:29",
      "content": "<p>This is my first deep learning competition so I have more questions than answers :) </p>\n\n<p>I think that the data preparation (same studio background, fix rotations) and the high resolution pixelwise loss function makes data augmentation less useful in here. Probably images taken in the wild would require more augmentation.</p>",
      "rawMarkdown": "This is my first deep learning competition so I have more questions than answers :) \n\nI think that the data preparation (same studio background, fix rotations) and the high resolution pixelwise loss function makes data augmentation less useful in here. Probably images taken in the wild would require more augmentation.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 218138,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "09/02/2017 13:51:21",
      "content": "<p>I'm just using flip. Have you quantified the effect of other augmentation methods?</p>",
      "votes": null,
      "replies": [
        {
          "id": 218158,
          "author_name": "gaborfodor",
          "author_url": "",
          "post_date": "09/02/2017 16:44:31",
          "content": "<p>Nothing conclusive yet. I am still playing with different network structures with lower resolution. My current training time is just too long for proper experiments.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 218171,
      "author_name": "markpopov",
      "author_url": "",
      "post_date": "09/02/2017 18:58:23",
      "content": "<p>Hi, beluga. I also use flip and it looks most effective so far, other augmentation methods makes result much worse. I am going to try some ideas from GANs, I have very strong overfitting. Do you think it could help to cope with overfitting?  </p>",
      "votes": null,
      "replies": [
        {
          "id": 218342,
          "author_name": "gaborfodor",
          "author_url": "",
          "post_date": "09/03/2017 20:30:29",
          "content": "<p>This is my first deep learning competition so I have more questions than answers :) </p>\n\n<p>I think that the data preparation (same studio background, fix rotations) and the high resolution pixelwise loss function makes data augmentation less useful in here. Probably images taken in the wild would require more augmentation.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "218011": "I collected a few methods. https://www.kaggle.com/gaborfodor/augmentation-methods",
    "218138": "I'm just using flip. Have you quantified the effect of other augmentation methods?",
    "218158": "Nothing conclusive yet. I am still playing with different network structures with lower resolution. My current training time is just too long for proper experiments.",
    "218171": "Hi, beluga. I also use flip and it looks most effective so far, other augmentation methods makes result much worse. I am going to try some ideas from GANs, I have very strong overfitting. Do you think it could help to cope with overfitting?",
    "218342": "This is my first deep learning competition so I have more questions than answers :) \n\nI think that the data preparation (same studio background, fix rotations) and the high resolution pixelwise loss function makes data augmentation less useful in here. Probably images taken in the wild would require more augmentation."
  },
  "source": "meta"
}