{
  "id": 98245,
  "title": "Finding dogs in the images first ?",
  "url": "/competitions/generative-dog-images/discussion/98245",
  "author_name": "",
  "post_date": "2019-07-02T09:09:09.901135Z",
  "votes": 1,
  "comment_count": 5,
  "views": 0,
  "content": "<p>In the dataset there are multiple images where other than dogs , other objects , human are there ? So , i have a general wondering , if there is a way to extract the dogs from the images and feed them to GAN , will there be better chances of getting good output ?</p>",
  "messages": [
    {
      "id": "566519",
      "postDate": "07/02/2019 09:09:09",
      "content": "<p>In the dataset there are multiple images where other than dogs , other objects , human are there ? So , i have a general wondering , if there is a way to extract the dogs from the images and feed them to GAN , will there be better chances of getting good output ?</p>",
      "rawMarkdown": "In the dataset there are multiple images where other than dogs , other objects , human are there ? So , i have a general wondering , if there is a way to extract the dogs from the images and feed them to GAN , will there be better chances of getting good output ?",
      "votes": null
    },
    {
      "id": "566590",
      "postDate": "07/02/2019 10:54:12",
      "content": "<p>Yes, you can e.g. crop the pictures using the bounding boxes.\nBut I am not 100% sure how that will affect private LB rating (magical model). I hope it will include some kind of pre trained network that only focusses on dogs and doesn't care about surroundings. </p>\n\n<p>There is already a kernel for that: <a href=\"https://www.kaggle.com/whizzkid/crop-images-using-bounding-box\">Link</a></p>",
      "rawMarkdown": "Yes, you can e.g. crop the pictures using the bounding boxes.\nBut I am not 100% sure how that will affect private LB rating (magical model). I hope it will include some kind of pre trained network that only focusses on dogs and doesn't care about surroundings. \n\nThere is already a kernel for that: [Link](https://www.kaggle.com/whizzkid/crop-images-using-bounding-box)",
      "votes": null
    },
    {
      "id": "566602",
      "postDate": "07/02/2019 11:02:59",
      "content": "<p>I wonder if the private dataset are portraits of dogs or dogs with surroundings. If private dataset has surroundings then we want to keep surroundings but if private is portraits then we want to remove surroundings and make portraits.</p>\n\n<p>This is the tricky thing about Kaggle competitions. In real life, you would have a better sense of what the private dataset contains.</p>",
      "rawMarkdown": "I wonder if the private dataset are portraits of dogs or dogs with surroundings. If private dataset has surroundings then we want to keep surroundings but if private is portraits then we want to remove surroundings and make portraits.\n\nThis is the tricky thing about Kaggle competitions. In real life, you would have a better sense of what the private dataset contains.",
      "votes": null
    },
    {
      "id": "566690",
      "postDate": "07/02/2019 13:15:28",
      "content": "<p>Thank you so much for taking your time to respond. This clears up</p>",
      "rawMarkdown": "Thank you so much for taking your time to respond. This clears up",
      "votes": null
    },
    {
      "id": "566748",
      "postDate": "07/02/2019 14:48:09",
      "content": "<p>Thank you for teaching me!</p>",
      "rawMarkdown": "Thank you for teaching me!",
      "votes": null
    },
    {
      "id": "572848",
      "postDate": "07/11/2019 13:19:42",
      "content": "<p><a href=\"/cdeotte\">@cdeotte</a> so we can select two kernels for final scoring, one with cropped dogs and the other one with full images, but there are so many ideas and different architectures of GANs so the choice will be difficult</p>",
      "rawMarkdown": "cdeotte so we can select two kernels for final scoring, one with cropped dogs and the other one with full images, but there are so many ideas and different architectures of GANs so the choice will be difficult",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 566590,
      "author_name": "ilu000",
      "author_url": "",
      "post_date": "07/02/2019 10:54:12",
      "content": "<p>Yes, you can e.g. crop the pictures using the bounding boxes.\nBut I am not 100% sure how that will affect private LB rating (magical model). I hope it will include some kind of pre trained network that only focusses on dogs and doesn't care about surroundings. </p>\n\n<p>There is already a kernel for that: <a href=\"https://www.kaggle.com/whizzkid/crop-images-using-bounding-box\">Link</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 566602,
          "author_name": "cdeotte",
          "author_url": "",
          "post_date": "07/02/2019 11:02:59",
          "content": "<p>I wonder if the private dataset are portraits of dogs or dogs with surroundings. If private dataset has surroundings then we want to keep surroundings but if private is portraits then we want to remove surroundings and make portraits.</p>\n\n<p>This is the tricky thing about Kaggle competitions. In real life, you would have a better sense of what the private dataset contains.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 566690,
          "author_name": "phoenix9032",
          "author_url": "",
          "post_date": "07/02/2019 13:15:28",
          "content": "<p>Thank you so much for taking your time to respond. This clears up</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 566748,
          "author_name": "snakayama",
          "author_url": "",
          "post_date": "07/02/2019 14:48:09",
          "content": "<p>Thank you for teaching me!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 572848,
          "author_name": "kostyaatarik",
          "author_url": "",
          "post_date": "07/11/2019 13:19:42",
          "content": "<p><a href=\"/cdeotte\">@cdeotte</a> so we can select two kernels for final scoring, one with cropped dogs and the other one with full images, but there are so many ideas and different architectures of GANs so the choice will be difficult</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "566519": "In the dataset there are multiple images where other than dogs , other objects , human are there ? So , i have a general wondering , if there is a way to extract the dogs from the images and feed them to GAN , will there be better chances of getting good output ?",
    "566590": "Yes, you can e.g. crop the pictures using the bounding boxes.\nBut I am not 100% sure how that will affect private LB rating (magical model). I hope it will include some kind of pre trained network that only focusses on dogs and doesn't care about surroundings. \n\nThere is already a kernel for that: [Link](https://www.kaggle.com/whizzkid/crop-images-using-bounding-box)",
    "566602": "I wonder if the private dataset are portraits of dogs or dogs with surroundings. If private dataset has surroundings then we want to keep surroundings but if private is portraits then we want to remove surroundings and make portraits.\n\nThis is the tricky thing about Kaggle competitions. In real life, you would have a better sense of what the private dataset contains.",
    "566690": "Thank you so much for taking your time to respond. This clears up",
    "566748": "Thank you for teaching me!",
    "572848": "cdeotte so we can select two kernels for final scoring, one with cropped dogs and the other one with full images, but there are so many ideas and different architectures of GANs so the choice will be difficult"
  },
  "source": "meta"
}