{
  "id": 39708,
  "title": "Predicted Masks - No Training",
  "url": "/competitions/carvana-image-masking-challenge/discussion/39708",
  "author_name": "",
  "post_date": "2017-09-19T18:55:03.076819600Z",
  "votes": 7,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Thought I would share this as I found it to be pretty impressive. Using a Unet that has never seen any training image is able to generate predicted masks that looks like this...</p>\n\n<p>I started to play with it more and it turns out that every time I re-load the model and run predictions again it generates a completely different looking mask as the model focuses on different aspects of the image each time...pretty cool!</p>\n\n<p>This is the power of deep learning :)</p>",
  "messages": [
    {
      "id": "222714",
      "postDate": "09/19/2017 18:55:03",
      "content": "<p>Thought I would share this as I found it to be pretty impressive. Using a Unet that has never seen any training image is able to generate predicted masks that looks like this...</p>\n\n<p>I started to play with it more and it turns out that every time I re-load the model and run predictions again it generates a completely different looking mask as the model focuses on different aspects of the image each time...pretty cool!</p>\n\n<p>This is the power of deep learning :)</p>",
      "rawMarkdown": "Thought I would share this as I found it to be pretty impressive. Using a Unet that has never seen any training image is able to generate predicted masks that looks like this...\n\nI started to play with it more and it turns out that every time I re-load the model and run predictions again it generates a completely different looking mask as the model focuses on different aspects of the image each time...pretty cool!\n\nThis is the power of deep learning :)",
      "votes": null
    },
    {
      "id": "222738",
      "postDate": "09/19/2017 20:16:49",
      "content": "<p>They seem to be edge detectors.</p>",
      "rawMarkdown": "They seem to be edge detectors.",
      "votes": null
    },
    {
      "id": "222771",
      "postDate": "09/20/2017 01:44:34",
      "content": "<p>Edges are inherently interesting to neural nets?</p>",
      "rawMarkdown": "Edges are inherently interesting to neural nets?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 222738,
      "author_name": "ironbar",
      "author_url": "",
      "post_date": "09/19/2017 20:16:49",
      "content": "<p>They seem to be edge detectors.</p>",
      "votes": null,
      "replies": [
        {
          "id": 222771,
          "author_name": "thewonz",
          "author_url": "",
          "post_date": "09/20/2017 01:44:34",
          "content": "<p>Edges are inherently interesting to neural nets?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "222714": "Thought I would share this as I found it to be pretty impressive. Using a Unet that has never seen any training image is able to generate predicted masks that looks like this...\n\nI started to play with it more and it turns out that every time I re-load the model and run predictions again it generates a completely different looking mask as the model focuses on different aspects of the image each time...pretty cool!\n\nThis is the power of deep learning :)",
    "222738": "They seem to be edge detectors.",
    "222771": "Edges are inherently interesting to neural nets?"
  },
  "source": "meta"
}