{
  "id": 17200,
  "title": "Welcome to the competition",
  "url": "/competitions/face-verification2/discussion/17200",
  "author_name": "",
  "post_date": "2015-10-28T07:50:09.263Z",
  "votes": 1,
  "comment_count": 1,
  "views": 696,
  "content": "<p>On behalf of the organizers, I'd like to welcome you to this face verification competition.</p>\n\n<p>The leaderboard has two benchmarks, of which the deep features distance benchmark is the more interesting one. The method simply computes the Euclidean distance between the 4096 dimensional deep feature vectors. The smaller the distance the more likely the two pictures are from the same person. As the last step, the distance is transformed to a probability by inverting the order and normalizing.</p>\n\n<p>Some points that may improve the result:</p>\n\n<ul>\n<li>The algorithm does not take into account the training data at all. So for example learning another distance metric instead of Euclidean may improve the accuracy.\n\n<ul><li>The AlexNet is not actually designed for this purpose, but just happens to work. Alternative nets may improve the performance.</li>\n<li>Look into the OpenFace framework: this should be very good for face recognition: <a href=\"https://github.com/cmusatyalab/openface\">https://github.com/cmusatyalab/openface</a></li></ul></li>\n</ul>\n\n<p>Corresponding Python implementation is attached. Matlab implementation will follow. Also other supporting code will be posted, so follow the forum and start testing your results.</p>",
  "messages": [
    {
      "id": "97543",
      "postDate": "10/28/2015 07:50:09",
      "content": "<p>On behalf of the organizers, I'd like to welcome you to this face verification competition.</p>\n\n<p>The leaderboard has two benchmarks, of which the deep features distance benchmark is the more interesting one. The method simply computes the Euclidean distance between the 4096 dimensional deep feature vectors. The smaller the distance the more likely the two pictures are from the same person. As the last step, the distance is transformed to a probability by inverting the order and normalizing.</p>\n\n<p>Some points that may improve the result:</p>\n\n<ul>\n<li>The algorithm does not take into account the training data at all. So for example learning another distance metric instead of Euclidean may improve the accuracy.\n\n<ul><li>The AlexNet is not actually designed for this purpose, but just happens to work. Alternative nets may improve the performance.</li>\n<li>Look into the OpenFace framework: this should be very good for face recognition: <a href=\"https://github.com/cmusatyalab/openface\">https://github.com/cmusatyalab/openface</a></li></ul></li>\n</ul>\n\n<p>Corresponding Python implementation is attached. Matlab implementation will follow. Also other supporting code will be posted, so follow the forum and start testing your results.</p>",
      "rawMarkdown": "On behalf of the organizers, I'd like to welcome you to this face verification competition.\r\n\r\nThe leaderboard has two benchmarks, of which the deep features distance benchmark is the more interesting one. The method simply computes the Euclidean distance between the 4096 dimensional deep feature vectors. The smaller the distance the more likely the two pictures are from the same person. As the last step, the distance is transformed to a probability by inverting the order and normalizing.\r\n\r\nSome points that may improve the result:\r\n\r\n - The algorithm does not take into account the training data at all. So for example learning another distance metric instead of Euclidean may improve the accuracy.\r\n- The AlexNet is not actually designed for this purpose, but just happens to work. Alternative nets may improve the performance.\r\n- Look into the OpenFace framework: this should be very good for face recognition: https://github.com/cmusatyalab/openface\r\n\r\nCorresponding Python implementation is attached. Matlab implementation will follow. Also other supporting code will be posted, so follow the forum and start testing your results.",
      "votes": null
    },
    {
      "id": "97972",
      "postDate": "11/02/2015 06:29:33",
      "content": "<p>Attached is also the Matlab version of the script.</p>",
      "rawMarkdown": "Attached is also the Matlab version of the script.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 97972,
      "author_name": "mahehu",
      "author_url": "",
      "post_date": "11/02/2015 06:29:33",
      "content": "<p>Attached is also the Matlab version of the script.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "97543": "On behalf of the organizers, I'd like to welcome you to this face verification competition.\r\n\r\nThe leaderboard has two benchmarks, of which the deep features distance benchmark is the more interesting one. The method simply computes the Euclidean distance between the 4096 dimensional deep feature vectors. The smaller the distance the more likely the two pictures are from the same person. As the last step, the distance is transformed to a probability by inverting the order and normalizing.\r\n\r\nSome points that may improve the result:\r\n\r\n - The algorithm does not take into account the training data at all. So for example learning another distance metric instead of Euclidean may improve the accuracy.\r\n- The AlexNet is not actually designed for this purpose, but just happens to work. Alternative nets may improve the performance.\r\n- Look into the OpenFace framework: this should be very good for face recognition: https://github.com/cmusatyalab/openface\r\n\r\nCorresponding Python implementation is attached. Matlab implementation will follow. Also other supporting code will be posted, so follow the forum and start testing your results.",
    "97972": "Attached is also the Matlab version of the script."
  },
  "source": "meta"
}