{
  "id": 12482,
  "title": "HOG+SVM",
  "url": "/competitions/datasciencebowl/discussion/12482",
  "author_name": "",
  "post_date": "2015-02-11T14:43:32.277Z",
  "votes": 4,
  "comment_count": 4,
  "views": 1618,
  "content": "<p>Does anyone use HOG transform+SVM? I tried, but I didn't find the results I hoped. It is worse than simple image manipulation+features selection.</p>",
  "messages": [
    {
      "id": "64031",
      "postDate": "02/11/2015 14:43:32",
      "content": "<p>Does anyone use HOG transform+SVM? I tried, but I didn't find the results I hoped. It is worse than simple image manipulation+features selection.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64163",
      "postDate": "02/13/2015 13:46:32",
      "content": "<p>When trying HOG+SVM, how are you handling the different image sizes to build the feature vector? &nbsp;Rescaling&nbsp;+&nbsp;padding, or just padding?</p>\n<p>Since some&nbsp;classes look alike but one is just bigger than the other, I would think it's best not to rescale them. &nbsp;So I was thinking of just putting the image in the center and then pad with white to the size of the largest. &nbsp;But since most of them are much smaller than the largest, I'd somehow have to give more weight to the center of the image.</p>\n<p>Another option would be to rescale to the same size in one dimension and pad in the other dimension, then also add the number of pixels in the original image as a feature.</p>\n<p>Also, how are you handling rotation? &nbsp;I guess when predicting you should try for each class a number of different rotations of each test image and select the one with the highest probability.</p>\n<p>Did you also look into PHOG, or just HOG?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "64173",
      "postDate": "02/13/2015 17:34:38",
      "content": "<p>Hi, I did not perform any padding, even to contain the number of features. I resized the images, but to stay with less than 500 features, the pixel for the inner cell cannot be too small.</p>\n<p>I tried just HOG.</p>\n<p>The problem for me is computational power to train SVM on a standard pc.</p>\n<p>So I tried to stay with fewer features.</p>\n<p>I obtained better results with Haralick features.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "445299",
      "postDate": "12/26/2018 07:00:57",
      "content": "<p>does HOG+SVM useful?   I didn't find example on KAGGLE!!</p>",
      "rawMarkdown": "does HOG+SVM useful?   I didn't find example on KAGGLE!!",
      "votes": null
    },
    {
      "id": "2328385",
      "postDate": "07/03/2023 14:56:27",
      "content": "<p>realtime HOG using GIF : <a href=\"https://youtu.be/Leg_GVDWcls\" target=\"_blank\">https://youtu.be/Leg_GVDWcls</a></p>",
      "rawMarkdown": "realtime HOG using GIF : https://youtu.be/Leg_GVDWcls",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2328385,
      "author_name": "bemorekgg",
      "author_url": "",
      "post_date": "07/03/2023 14:56:27",
      "content": "<p>realtime HOG using GIF : <a href=\"https://youtu.be/Leg_GVDWcls\" target=\"_blank\">https://youtu.be/Leg_GVDWcls</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64163,
      "author_name": "hermanbovens",
      "author_url": "",
      "post_date": "02/13/2015 13:46:32",
      "content": "<p>When trying HOG+SVM, how are you handling the different image sizes to build the feature vector? &nbsp;Rescaling&nbsp;+&nbsp;padding, or just padding?</p>\n<p>Since some&nbsp;classes look alike but one is just bigger than the other, I would think it's best not to rescale them. &nbsp;So I was thinking of just putting the image in the center and then pad with white to the size of the largest. &nbsp;But since most of them are much smaller than the largest, I'd somehow have to give more weight to the center of the image.</p>\n<p>Another option would be to rescale to the same size in one dimension and pad in the other dimension, then also add the number of pixels in the original image as a feature.</p>\n<p>Also, how are you handling rotation? &nbsp;I guess when predicting you should try for each class a number of different rotations of each test image and select the one with the highest probability.</p>\n<p>Did you also look into PHOG, or just HOG?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 64173,
      "author_name": "elenacuoco",
      "author_url": "",
      "post_date": "02/13/2015 17:34:38",
      "content": "<p>Hi, I did not perform any padding, even to contain the number of features. I resized the images, but to stay with less than 500 features, the pixel for the inner cell cannot be too small.</p>\n<p>I tried just HOG.</p>\n<p>The problem for me is computational power to train SVM on a standard pc.</p>\n<p>So I tried to stay with fewer features.</p>\n<p>I obtained better results with Haralick features.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 445299,
      "author_name": "nckustudent",
      "author_url": "",
      "post_date": "12/26/2018 07:00:57",
      "content": "<p>does HOG+SVM useful?   I didn't find example on KAGGLE!!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "64031": "",
    "64163": "",
    "64173": "",
    "445299": "does HOG+SVM useful?   I didn't find example on KAGGLE!!",
    "2328385": "realtime HOG using GIF : https://youtu.be/Leg_GVDWcls"
  },
  "source": "meta"
}