{
  "id": 17753,
  "title": "Classifiers",
  "url": "/competitions/noaa-right-whale-recognition/discussion/17753",
  "author_name": "George Fisher",
  "post_date": "2015-12-07T16:09:46.987000",
  "votes": 0,
  "comment_count": 0,
  "views": 446,
  "content": "<p>I have spent a good deal of time training classifiers: OpenCV 3.0.0 LBP and HAAR,  DLib 18.17 HOG and  MATLAB R2015b.      </p>\n\n<p>The LBP and HAAR classifiers produced a 50-65% accuracy, defined as the ability to correctly identify a small-ish bounding box around the identifiable part of the whales in the test set. Since these classifiers produce more than one bounding box, I chose the one with the largest area from the process of running the classifiers on the three color channels of the RGB images.   </p>\n\n<p>The HOG classifier succeeded in surrounding the whales some 95% of the time but the bounding box area was barely smaller than that of the original image. The MATLAB classifier produced the opposite result: a blizzard of little boxes that provided no way of distinguishing the best one that I could cook up.   </p>\n\n<p>I have given up on this approach but I have attached the results of my efforts in case anyone else can make better use of them than I did.      </p>\n\n<p>I am starting to have some success with a convolutional neural network built with nolearn/lasagne/theano, but so-far 'success' only in a relative sense. I'll post any real success I have with this if I can manage such a thing before time runs out.   </p>",
  "messages": [
    {
      "id": 100449,
      "postDate": "2015-12-07T16:09:46.987Z",
      "content": "<p>I have spent a good deal of time training classifiers: OpenCV 3.0.0 LBP and HAAR,  DLib 18.17 HOG and  MATLAB R2015b.      </p>\n\n<p>The LBP and HAAR classifiers produced a 50-65% accuracy, defined as the ability to correctly identify a small-ish bounding box around the identifiable part of the whales in the test set. Since these classifiers produce more than one bounding box, I chose the one with the largest area from the process of running the classifiers on the three color channels of the RGB images.   </p>\n\n<p>The HOG classifier succeeded in surrounding the whales some 95% of the time but the bounding box area was barely smaller than that of the original image. The MATLAB classifier produced the opposite result: a blizzard of little boxes that provided no way of distinguishing the best one that I could cook up.   </p>\n\n<p>I have given up on this approach but I have attached the results of my efforts in case anyone else can make better use of them than I did.      </p>\n\n<p>I am starting to have some success with a convolutional neural network built with nolearn/lasagne/theano, but so-far 'success' only in a relative sense. I'll post any real success I have with this if I can manage such a thing before time runs out.   </p>",
      "rawMarkdown": "I have spent a good deal of time training classifiers: OpenCV 3.0.0 LBP and HAAR,  DLib 18.17 HOG and  MATLAB R2015b.      \r\n\r\nThe LBP and HAAR classifiers produced a 50-65% accuracy, defined as the ability to correctly identify a small-ish bounding box around the identifiable part of the whales in the test set. Since these classifiers produce more than one bounding box, I chose the one with the largest area from the process of running the classifiers on the three color channels of the RGB images.   \r\n\r\nThe HOG classifier succeeded in surrounding the whales some 95% of the time but the bounding box area was barely smaller than that of the original image. The MATLAB classifier produced the opposite result: a blizzard of little boxes that provided no way of distinguishing the best one that I could cook up.   \r\n\r\nI have given up on this approach but I have attached the results of my efforts in case anyone else can make better use of them than I did.      \r\n\r\nI am starting to have some success with a convolutional neural network built with nolearn/lasagne/theano, but so-far 'success' only in a relative sense. I'll post any real success I have with this if I can manage such a thing before time runs out.   "
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "100449": "I have spent a good deal of time training classifiers: OpenCV 3.0.0 LBP and HAAR,  DLib 18.17 HOG and  MATLAB R2015b.      \r\n\r\nThe LBP and HAAR classifiers produced a 50-65% accuracy, defined as the ability to correctly identify a small-ish bounding box around the identifiable part of the whales in the test set. Since these classifiers produce more than one bounding box, I chose the one with the largest area from the process of running the classifiers on the three color channels of the RGB images.   \r\n\r\nThe HOG classifier succeeded in surrounding the whales some 95% of the time but the bounding box area was barely smaller than that of the original image. The MATLAB classifier produced the opposite result: a blizzard of little boxes that provided no way of distinguishing the best one that I could cook up.   \r\n\r\nI have given up on this approach but I have attached the results of my efforts in case anyone else can make better use of them than I did.      \r\n\r\nI am starting to have some success with a convolutional neural network built with nolearn/lasagne/theano, but so-far 'success' only in a relative sense. I'll post any real success I have with this if I can manage such a thing before time runs out.   "
  }
}