{
  "id": 67160,
  "title": "Welcome the Delay",
  "url": "/competitions/airbus-ship-detection/discussion/67160",
  "author_name": "Peter",
  "post_date": "2018-09-29T12:31:11.463000",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I noticed that due to the delay of this competition, I took step back and spend some time better understanding the real problem at hand.</p>\n\n<p>Rather then diving straight into some flavour of a CNN and go through many “try-run-tune” iterations, I instead read some interesting articles and white papers on the challenges of object detection within the context of satellite images.</p>\n\n<p>And as a result, I’m a bit wiser today than I would have been otherwise. Although I love the competition element of Kaggle, it sometimes results in being caught up too much in the moment.</p>\n\n<p>So thanks for the delay!</p>",
  "messages": [
    {
      "id": 395803,
      "postDate": "2018-09-29T12:31:11.463Z",
      "content": "<p>I noticed that due to the delay of this competition, I took step back and spend some time better understanding the real problem at hand.</p>\n\n<p>Rather then diving straight into some flavour of a CNN and go through many “try-run-tune” iterations, I instead read some interesting articles and white papers on the challenges of object detection within the context of satellite images.</p>\n\n<p>And as a result, I’m a bit wiser today than I would have been otherwise. Although I love the competition element of Kaggle, it sometimes results in being caught up too much in the moment.</p>\n\n<p>So thanks for the delay!</p>",
      "rawMarkdown": "I noticed that due to the delay of this competition, I took step back and spend some time better understanding the real problem at hand.\n\nRather then diving straight into some flavour of a CNN and go through many “try-run-tune” iterations, I instead read some interesting articles and white papers on the challenges of object detection within the context of satellite images.\n\nAnd as a result, I’m a bit wiser today than I would have been otherwise. Although I love the competition element of Kaggle, it sometimes results in being caught up too much in the moment.\n\nSo thanks for the delay!\n",
      "votes": 3
    },
    {
      "id": 396390,
      "postDate": "2018-09-30T16:20:53.400Z",
      "rawMarkdown": "",
      "votes": 1,
      "isDeleted": true,
      "replies": [
        {
          "id": 396414,
          "postDate": "2018-09-30T17:35:49.057Z",
          "content": "<p>Sure, one of the challenges is that the algorithm should be able to support fine grained object location but also needs to be context aware.</p>\n\n<p>So if looking at some possible approaches:</p>\n\n<ul>\n<li><p>Region based CNN solutions (like fast rcnn) are good at locating smaller objects. But\nthey are not context aware, and as a result they tend to wrongly locate ships\non land and other “impossible” areas.</p></li>\n<li><p>SSD/YOLO type of CNN solutions are context aware and can better identify and\nexclude land areas. However they are not able to differentiate ships close to each other due to the course grained grid they are typically using.</p></li>\n</ul>\n\n<p>So have to find a solution that marries best of both worlds. </p>",
          "rawMarkdown": "Sure, one of the challenges is that the algorithm should be able to support fine grained object location but also needs to be context aware.\n\nSo if looking at some possible approaches:\n\n - Region based CNN solutions (like fast rcnn) are good at locating smaller objects. But\n   they are not context aware, and as a result they tend to wrongly locate ships\n   on land and other “impossible” areas.\n   \n - SSD/YOLO type of CNN solutions are context aware and can better identify and\n   exclude land areas. However they are not able to differentiate ships close to each other due to the course grained grid they are typically using.\n\nSo have to find a solution that marries best of both worlds. \n",
          "votes": 3
        }
      ]
    }
  ],
  "comments": [
    {
      "id": 396390,
      "author_name": "",
      "author_url": "",
      "post_date": "2018-09-30T16:20:53.400000",
      "content": "",
      "votes": 1,
      "replies": [
        {
          "id": 396414,
          "author_name": "Peter",
          "author_url": "",
          "post_date": "2018-09-30T17:35:49.057000",
          "content": "<p>Sure, one of the challenges is that the algorithm should be able to support fine grained object location but also needs to be context aware.</p>\n\n<p>So if looking at some possible approaches:</p>\n\n<ul>\n<li><p>Region based CNN solutions (like fast rcnn) are good at locating smaller objects. But\nthey are not context aware, and as a result they tend to wrongly locate ships\non land and other “impossible” areas.</p></li>\n<li><p>SSD/YOLO type of CNN solutions are context aware and can better identify and\nexclude land areas. However they are not able to differentiate ships close to each other due to the course grained grid they are typically using.</p></li>\n</ul>\n\n<p>So have to find a solution that marries best of both worlds. </p>",
          "votes": 3,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "395803": "I noticed that due to the delay of this competition, I took step back and spend some time better understanding the real problem at hand.\n\nRather then diving straight into some flavour of a CNN and go through many “try-run-tune” iterations, I instead read some interesting articles and white papers on the challenges of object detection within the context of satellite images.\n\nAnd as a result, I’m a bit wiser today than I would have been otherwise. Although I love the competition element of Kaggle, it sometimes results in being caught up too much in the moment.\n\nSo thanks for the delay!\n",
    "396390": ""
  }
}