{
  "id": 181392,
  "title": "A trick to improve result",
  "url": "/competitions/landmark-recognition-2020/discussion/181392",
  "author_name": "",
  "post_date": "2020-09-08T16:54:25.554639700Z",
  "votes": 8,
  "comment_count": 2,
  "views": 0,
  "content": "<ul>\n<li>Convert images to vectors using pre-trained network</li>\n<li>For each image from test set they found K closest images from train set(K=5)</li>\n<li>For each landmark_id we computed: scores[landmark_id] = sum(cos(query_image,index)for index in K_closest images)</li>\n<li>For each landmark_id we normalized its score: scores[landmark_id]/=min(K,number of  samples in train dataset with landmark = landmark_id)</li>\n<li>label = argmax(scores), confidence = scores[label]</li>\n</ul>",
  "messages": [
    {
      "id": "1003096",
      "postDate": "09/08/2020 16:54:25",
      "content": "<ul>\n<li>Convert images to vectors using pre-trained network</li>\n<li>For each image from test set they found K closest images from train set(K=5)</li>\n<li>For each landmark_id we computed: scores[landmark_id] = sum(cos(query_image,index)for index in K_closest images)</li>\n<li>For each landmark_id we normalized its score: scores[landmark_id]/=min(K,number of  samples in train dataset with landmark = landmark_id)</li>\n<li>label = argmax(scores), confidence = scores[label]</li>\n</ul>",
      "rawMarkdown": "Convert images to vectors using pre-trained network\n- For each image from test set they found K closest images from train set(K=5)\n- For each landmark_id we computed: scores[landmark_id] = sum(cos(query_image,index)for index in K_closest images)\n- For each landmark_id we normalized its score: scores[landmark_id]/=min(K,number of  samples in train dataset with landmark = landmark_id)\n- label = argmax(scores), confidence = scores[label]",
      "votes": null
    },
    {
      "id": "1005410",
      "postDate": "09/10/2020 13:29:00",
      "content": "<p>and where is the trick?</p>",
      "rawMarkdown": "and where is the trick?",
      "votes": null
    },
    {
      "id": "1005421",
      "postDate": "09/10/2020 13:39:18",
      "content": "<p><a href=\"https://www.kaggle.com/christofhenkel\" target=\"_blank\">@christofhenkel</a>  ☝️</p>",
      "rawMarkdown": "christofhenkel  ☝️",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1005410,
      "author_name": "christofhenkel",
      "author_url": "",
      "post_date": "09/10/2020 13:29:00",
      "content": "<p>and where is the trick?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1005421,
      "author_name": "penchalaiah123",
      "author_url": "",
      "post_date": "09/10/2020 13:39:18",
      "content": "<p><a href=\"https://www.kaggle.com/christofhenkel\" target=\"_blank\">@christofhenkel</a>  ☝️</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1003096": "Convert images to vectors using pre-trained network\n- For each image from test set they found K closest images from train set(K=5)\n- For each landmark_id we computed: scores[landmark_id] = sum(cos(query_image,index)for index in K_closest images)\n- For each landmark_id we normalized its score: scores[landmark_id]/=min(K,number of  samples in train dataset with landmark = landmark_id)\n- label = argmax(scores), confidence = scores[label]",
    "1005410": "and where is the trick?",
    "1005421": "christofhenkel  ☝️"
  },
  "source": "meta"
}