{
  "id": 38007,
  "title": "Car color distribution. Blue car advantage?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/38007",
  "author_name": "",
  "post_date": "2017-08-13T13:40:27.943124200Z",
  "votes": 5,
  "comment_count": 1,
  "views": 0,
  "content": "<p>P.S. Sorry, I wasn't able to publish kernel on kernel section. So I uploaded it to gitlab. </p>\n\n<p>Kernel itself - <a href=\"https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html\">https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html</a></p>\n\n<p>Plots from kernel:</p>\n\n<p><strong>Car color distribution in train and test:</strong>\n<img src=\"https://i.imgur.com/Xq6fttb.png\" alt=\"\" title=\"\"></p>\n\n<p><strong>Mean dice score by color:</strong>\n<img src=\"https://i.imgur.com/9kWTcEJ.png\" alt=\"\" title=\"\"></p>\n\n<p>It's a shame. Roayal Blue has the highest score but barely present in test!</p>\n\n<p>Seems like black is not so good. I guess it's because shadows under the car.</p>\n\n<p>Light gray? Try to avoid it)</p>\n\n<p>So the hypotesys, that colors, that are quite similar to backgrounds, perform worse than the other colors seems to be legit</p>\n\n<p>Any ideas how to preproccess these light colors?</p>\n\n<p>P.P.S. I added car colors csv. </p>",
  "messages": [
    {
      "id": "213028",
      "postDate": "08/13/2017 13:40:27",
      "content": "<p>P.S. Sorry, I wasn't able to publish kernel on kernel section. So I uploaded it to gitlab. </p>\n\n<p>Kernel itself - <a href=\"https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html\">https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html</a></p>\n\n<p>Plots from kernel:</p>\n\n<p><strong>Car color distribution in train and test:</strong>\n<img src=\"https://i.imgur.com/Xq6fttb.png\" alt=\"\" title=\"\"></p>\n\n<p><strong>Mean dice score by color:</strong>\n<img src=\"https://i.imgur.com/9kWTcEJ.png\" alt=\"\" title=\"\"></p>\n\n<p>It's a shame. Roayal Blue has the highest score but barely present in test!</p>\n\n<p>Seems like black is not so good. I guess it's because shadows under the car.</p>\n\n<p>Light gray? Try to avoid it)</p>\n\n<p>So the hypotesys, that colors, that are quite similar to backgrounds, perform worse than the other colors seems to be legit</p>\n\n<p>Any ideas how to preproccess these light colors?</p>\n\n<p>P.P.S. I added car colors csv. </p>",
      "rawMarkdown": "P.S. Sorry, I wasn't able to publish kernel on kernel section. So I uploaded it to gitlab. \n\nKernel itself - https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html\n\nPlots from kernel:\n\n**Car color distribution in train and test:**\n![][1]\n\n**Mean dice score by color:**\n![][2]\n\n\n  [1]: https://i.imgur.com/Xq6fttb.png\n  [2]: https://i.imgur.com/9kWTcEJ.png\n\n\nIt's a shame. Roayal Blue has the highest score but barely present in test!\n\nSeems like black is not so good. I guess it's because shadows under the car.\n\nLight gray? Try to avoid it)\n\nSo the hypotesys, that colors, that are quite similar to backgrounds, perform worse than the other colors seems to be legit\n\nAny ideas how to preproccess these light colors?\n\nP.P.S. I added car colors csv.",
      "votes": null
    },
    {
      "id": "213070",
      "postDate": "08/13/2017 16:21:22",
      "content": "<p>What if you weight the mean dice score by number of instances of each color? The variance will matter depending on how many cars of each color there are. </p>",
      "rawMarkdown": "What if you weight the mean dice score by number of instances of each color? The variance will matter depending on how many cars of each color there are.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 213070,
      "author_name": "craigglastonbury",
      "author_url": "",
      "post_date": "08/13/2017 16:21:22",
      "content": "<p>What if you weight the mean dice score by number of instances of each color? The variance will matter depending on how many cars of each color there are. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "213028": "P.S. Sorry, I wasn't able to publish kernel on kernel section. So I uploaded it to gitlab. \n\nKernel itself - https://feels_g00d_man.gitlab.io/plain-html/Kaggle/carvana_car_color.html\n\nPlots from kernel:\n\n**Car color distribution in train and test:**\n![][1]\n\n**Mean dice score by color:**\n![][2]\n\n\n  [1]: https://i.imgur.com/Xq6fttb.png\n  [2]: https://i.imgur.com/9kWTcEJ.png\n\n\nIt's a shame. Roayal Blue has the highest score but barely present in test!\n\nSeems like black is not so good. I guess it's because shadows under the car.\n\nLight gray? Try to avoid it)\n\nSo the hypotesys, that colors, that are quite similar to backgrounds, perform worse than the other colors seems to be legit\n\nAny ideas how to preproccess these light colors?\n\nP.P.S. I added car colors csv.",
    "213070": "What if you weight the mean dice score by number of instances of each color? The variance will matter depending on how many cars of each color there are."
  },
  "source": "meta"
}