{
  "id": 20554,
  "title": "Distance's Matrix",
  "url": "/competitions/draper-satellite-image-chronology/discussion/20554",
  "author_name": "Rodolphe Lampe",
  "post_date": "2016-04-29T20:22:06.613000",
  "votes": 2,
  "comment_count": 0,
  "views": 531,
  "content": "<p>A few ideas :\nFor a set of 5 images image_1, ..., image_5, we can, for every pair (image_i, image_j) : </p>\n\n<ul>\n<li>find the intersection and rotate appropriately (using neural nets)</li>\n<li>compute the number of different objects</li>\n<li>set a_{i,j} = this number</li>\n</ul>\n\n<p>Then use this matrix to find the order that minizes the distance between consecutive images. The problem with this approach is that it doesn't distinguish between the symetric orderings like 1 2 3 4 5 and 5 4 3 2 1</p>\n\n<p>We also have to learn how to recognize the car driving on a road, so that we don't count them for the number of different objects.</p>\n\n<hr>\n\n<p>The huge parkings can help to find order : there are definitely patterns to identify time.</p>\n\n<hr>\n\n<p>Entropy is increasing over time. I don't think if it will help but to compare image of forest, plain and nature in general, we could analyze how entropy is increasing (or not) over time ... The window of time is probably too small.\nWikipedia : Entropy is the only quantity in the physical sciences (apart from certain rare interactions in particle physics; see below) that requires a particular direction for time, sometimes called an arrow of time. As one goes &quot;forward&quot; in time, the second law of thermodynamics says, the entropy of an isolated system can increase, but not decrease. Hence, from one perspective, entropy measurement is a way of distinguishing the past from the future.</p>\n\n<hr>\n\n<p>Take the intersection of the 5 images. Analyze how brightness, constrast, etc are changing over time. Is there any pattern ?</p>\n\n<hr>\n\n<p>There are a lot of residential neighboorhood with lots of car. We could analyze how the car are parking : are they parking behind other cars moving forward or are they more likely to park moving backward ? I think it's not evenly distributed, specially if you consider the density of available places in the street. If we know how it's distributed, we can compare images to establish a probability that one image is before or after an other one.</p>\n\n<hr>\n\n<p>Teach a nnet to recognize between the few different situations (or do it by hand) : huge parkings, houses, nature, containers. Then for each situations, find the features that can identify time.</p>\n\n<p>I have never done computer vision. If someone wants to team up, I can help for everything else. I think there is a lot to do that's not necessarily computer vision related.</p>",
  "messages": [
    {
      "id": 117608,
      "postDate": "2016-04-29T20:22:06.613Z",
      "content": "<p>A few ideas :\nFor a set of 5 images image_1, ..., image_5, we can, for every pair (image_i, image_j) : </p>\n\n<ul>\n<li>find the intersection and rotate appropriately (using neural nets)</li>\n<li>compute the number of different objects</li>\n<li>set a_{i,j} = this number</li>\n</ul>\n\n<p>Then use this matrix to find the order that minizes the distance between consecutive images. The problem with this approach is that it doesn't distinguish between the symetric orderings like 1 2 3 4 5 and 5 4 3 2 1</p>\n\n<p>We also have to learn how to recognize the car driving on a road, so that we don't count them for the number of different objects.</p>\n\n<hr>\n\n<p>The huge parkings can help to find order : there are definitely patterns to identify time.</p>\n\n<hr>\n\n<p>Entropy is increasing over time. I don't think if it will help but to compare image of forest, plain and nature in general, we could analyze how entropy is increasing (or not) over time ... The window of time is probably too small.\nWikipedia : Entropy is the only quantity in the physical sciences (apart from certain rare interactions in particle physics; see below) that requires a particular direction for time, sometimes called an arrow of time. As one goes &quot;forward&quot; in time, the second law of thermodynamics says, the entropy of an isolated system can increase, but not decrease. Hence, from one perspective, entropy measurement is a way of distinguishing the past from the future.</p>\n\n<hr>\n\n<p>Take the intersection of the 5 images. Analyze how brightness, constrast, etc are changing over time. Is there any pattern ?</p>\n\n<hr>\n\n<p>There are a lot of residential neighboorhood with lots of car. We could analyze how the car are parking : are they parking behind other cars moving forward or are they more likely to park moving backward ? I think it's not evenly distributed, specially if you consider the density of available places in the street. If we know how it's distributed, we can compare images to establish a probability that one image is before or after an other one.</p>\n\n<hr>\n\n<p>Teach a nnet to recognize between the few different situations (or do it by hand) : huge parkings, houses, nature, containers. Then for each situations, find the features that can identify time.</p>\n\n<p>I have never done computer vision. If someone wants to team up, I can help for everything else. I think there is a lot to do that's not necessarily computer vision related.</p>",
      "rawMarkdown": "A few ideas :\r\nFor a set of 5 images image_1, ..., image_5, we can, for every pair (image_i, image_j) : \r\n\r\n- find the intersection and rotate appropriately (using neural nets)\r\n- compute the number of different objects\r\n- set a_{i,j} = this number\r\n\r\nThen use this matrix to find the order that minizes the distance between consecutive images. The problem with this approach is that it doesn't distinguish between the symetric orderings like 1 2 3 4 5 and 5 4 3 2 1\r\n\r\nWe also have to learn how to recognize the car driving on a road, so that we don't count them for the number of different objects.\r\n__________________________\r\nThe huge parkings can help to find order : there are definitely patterns to identify time.\r\n___________________________\r\nEntropy is increasing over time. I don't think if it will help but to compare image of forest, plain and nature in general, we could analyze how entropy is increasing (or not) over time ... The window of time is probably too small.\r\nWikipedia : Entropy is the only quantity in the physical sciences (apart from certain rare interactions in particle physics; see below) that requires a particular direction for time, sometimes called an arrow of time. As one goes \"forward\" in time, the second law of thermodynamics says, the entropy of an isolated system can increase, but not decrease. Hence, from one perspective, entropy measurement is a way of distinguishing the past from the future.\r\n___________________________\r\nTake the intersection of the 5 images. Analyze how brightness, constrast, etc are changing over time. Is there any pattern ?\r\n___________________________\r\nThere are a lot of residential neighboorhood with lots of car. We could analyze how the car are parking : are they parking behind other cars moving forward or are they more likely to park moving backward ? I think it's not evenly distributed, specially if you consider the density of available places in the street. If we know how it's distributed, we can compare images to establish a probability that one image is before or after an other one.\r\n___________________________\r\nTeach a nnet to recognize between the few different situations (or do it by hand) : huge parkings, houses, nature, containers. Then for each situations, find the features that can identify time.\r\n\r\n\r\nI have never done computer vision. If someone wants to team up, I can help for everything else. I think there is a lot to do that's not necessarily computer vision related.",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "117608": "A few ideas :\r\nFor a set of 5 images image_1, ..., image_5, we can, for every pair (image_i, image_j) : \r\n\r\n- find the intersection and rotate appropriately (using neural nets)\r\n- compute the number of different objects\r\n- set a_{i,j} = this number\r\n\r\nThen use this matrix to find the order that minizes the distance between consecutive images. The problem with this approach is that it doesn't distinguish between the symetric orderings like 1 2 3 4 5 and 5 4 3 2 1\r\n\r\nWe also have to learn how to recognize the car driving on a road, so that we don't count them for the number of different objects.\r\n__________________________\r\nThe huge parkings can help to find order : there are definitely patterns to identify time.\r\n___________________________\r\nEntropy is increasing over time. I don't think if it will help but to compare image of forest, plain and nature in general, we could analyze how entropy is increasing (or not) over time ... The window of time is probably too small.\r\nWikipedia : Entropy is the only quantity in the physical sciences (apart from certain rare interactions in particle physics; see below) that requires a particular direction for time, sometimes called an arrow of time. As one goes \"forward\" in time, the second law of thermodynamics says, the entropy of an isolated system can increase, but not decrease. Hence, from one perspective, entropy measurement is a way of distinguishing the past from the future.\r\n___________________________\r\nTake the intersection of the 5 images. Analyze how brightness, constrast, etc are changing over time. Is there any pattern ?\r\n___________________________\r\nThere are a lot of residential neighboorhood with lots of car. We could analyze how the car are parking : are they parking behind other cars moving forward or are they more likely to park moving backward ? I think it's not evenly distributed, specially if you consider the density of available places in the street. If we know how it's distributed, we can compare images to establish a probability that one image is before or after an other one.\r\n___________________________\r\nTeach a nnet to recognize between the few different situations (or do it by hand) : huge parkings, houses, nature, containers. Then for each situations, find the features that can identify time.\r\n\r\n\r\nI have never done computer vision. If someone wants to team up, I can help for everything else. I think there is a lot to do that's not necessarily computer vision related."
  }
}