{
  "id": 37659,
  "title": "Glasses",
  "url": "/competitions/carvana-image-masking-challenge/discussion/37659",
  "author_name": "",
  "post_date": "2017-08-07T01:04:18.249787700Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>UPDATE: I tried to delete the post but this is not possible, so instead I'll try to remember what I wrote originally to not waste time from people who open the thread.</p>\n\n<p>From the problem description, we can see that Carvana wants masks to be able to <strong>compose images with new and different backgrounds</strong>, among other things.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-competitions/kaggle/3333/media/carvana_graphics.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>In the above example, I don't think simply setting transparency to the green area and adding a background layer would result in a quality result because of the transparent areas (glasses) and in case of other angles reflective materials (mirrors, etc).</p>\n\n<p>I'm wondering whether there's any plan on how to handle that (as post-processing) to make models generated by  competition useful to generate realistic compositions, or something else I'm missing.</p>\n\n<p>Alternatively, I think the problem could be defined as a Alpha Matting, with proper masks considering transparent surfaces and that would enable much higher results for the company and challenge for us here :-) Of course the task of generating the grouth truth would be much more time consuming, not sure that would be viable.</p>",
  "messages": [
    {
      "id": "210761",
      "postDate": "08/07/2017 01:04:18",
      "content": "<p>UPDATE: I tried to delete the post but this is not possible, so instead I'll try to remember what I wrote originally to not waste time from people who open the thread.</p>\n\n<p>From the problem description, we can see that Carvana wants masks to be able to <strong>compose images with new and different backgrounds</strong>, among other things.</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-competitions/kaggle/3333/media/carvana_graphics.png\" alt=\"enter image description here\" title=\"\"></p>\n\n<p>In the above example, I don't think simply setting transparency to the green area and adding a background layer would result in a quality result because of the transparent areas (glasses) and in case of other angles reflective materials (mirrors, etc).</p>\n\n<p>I'm wondering whether there's any plan on how to handle that (as post-processing) to make models generated by  competition useful to generate realistic compositions, or something else I'm missing.</p>\n\n<p>Alternatively, I think the problem could be defined as a Alpha Matting, with proper masks considering transparent surfaces and that would enable much higher results for the company and challenge for us here :-) Of course the task of generating the grouth truth would be much more time consuming, not sure that would be viable.</p>",
      "rawMarkdown": "UPDATE: I tried to delete the post but this is not possible, so instead I'll try to remember what I wrote originally to not waste time from people who open the thread.\n\nFrom the problem description, we can see that Carvana wants masks to be able to **compose images with new and different backgrounds**, among other things.\n\n![enter image description here][1]\n\nIn the above example, I don't think simply setting transparency to the green area and adding a background layer would result in a quality result because of the transparent areas (glasses) and in case of other angles reflective materials (mirrors, etc).\n\nI'm wondering whether there's any plan on how to handle that (as post-processing) to make models generated by  competition useful to generate realistic compositions, or something else I'm missing.\n\nAlternatively, I think the problem could be defined as a Alpha Matting, with proper masks considering transparent surfaces and that would enable much higher results for the company and challenge for us here :-) Of course the task of generating the grouth truth would be much more time consuming, not sure that would be viable.\n\n  [1]: https://kaggle2.blob.core.windows.net/competitions/kaggle/3333/media/carvana_graphics.png",
      "votes": null
    },
    {
      "id": "211550",
      "postDate": "08/09/2017 10:44:22",
      "content": "<p>There are no plans to implement environment mapping / ray tracing (for reflective surfaces) or semi-transparency (for windows). I agree the result would be more photo-realistic with more diverse environments, but as described, the competition will yield assets that work very well in many situations.</p>",
      "rawMarkdown": "There are no plans to implement environment mapping / ray tracing (for reflective surfaces) or semi-transparency (for windows). I agree the result would be more photo-realistic with more diverse environments, but as described, the competition will yield assets that work very well in many situations.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 211550,
      "author_name": "brianshaler",
      "author_url": "",
      "post_date": "08/09/2017 10:44:22",
      "content": "<p>There are no plans to implement environment mapping / ray tracing (for reflective surfaces) or semi-transparency (for windows). I agree the result would be more photo-realistic with more diverse environments, but as described, the competition will yield assets that work very well in many situations.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "210761": "UPDATE: I tried to delete the post but this is not possible, so instead I'll try to remember what I wrote originally to not waste time from people who open the thread.\n\nFrom the problem description, we can see that Carvana wants masks to be able to **compose images with new and different backgrounds**, among other things.\n\n![enter image description here][1]\n\nIn the above example, I don't think simply setting transparency to the green area and adding a background layer would result in a quality result because of the transparent areas (glasses) and in case of other angles reflective materials (mirrors, etc).\n\nI'm wondering whether there's any plan on how to handle that (as post-processing) to make models generated by  competition useful to generate realistic compositions, or something else I'm missing.\n\nAlternatively, I think the problem could be defined as a Alpha Matting, with proper masks considering transparent surfaces and that would enable much higher results for the company and challenge for us here :-) Of course the task of generating the grouth truth would be much more time consuming, not sure that would be viable.\n\n  [1]: https://kaggle2.blob.core.windows.net/competitions/kaggle/3333/media/carvana_graphics.png",
    "211550": "There are no plans to implement environment mapping / ray tracing (for reflective surfaces) or semi-transparency (for windows). I agree the result would be more photo-realistic with more diverse environments, but as described, the competition will yield assets that work very well in many situations."
  },
  "source": "meta"
}