{
  "id": 187720,
  "title": "My engineering approach to the problem. #49",
  "url": "/competitions/landmark-recognition-2020/discussion/187720",
  "author_name": "syam puranam",
  "post_date": "2020-09-30T03:13:54.307000",
  "votes": 5,
  "comment_count": 0,
  "views": 0,
  "content": "<p>This is my first real competition. Definitely learnt quite a bit! Thanks to the organizers. After giving up on training my model, I leveraged the baseline model with few additional measures. My approach did quite well in my test cases with 2019 competition data (~0.6) but unfortunately didn't do quite as well on this contest.</p>\n<ul>\n<li><p>Precomputed embeddings for complete train data and stored in sqlite3 database. This substantially reduced time to evaluate.</p></li>\n<li><p>Mix in precomputed embeddings with embeddings generated on a subset openimages dataset. I assign a landmark ID of 999999  to this dataset and use it to filter out any images that matched this dataset. </p></li>\n<li><p>After this step, the dataset is reduced by 10x allowing a lot of time to do local matching etc. </p></li>\n</ul>\n<p>I did try using cv2 sift for local scoring but didn't see much benefit. </p>",
  "messages": [
    {
      "id": 1032186,
      "postDate": "2020-09-30T03:13:54.307Z",
      "content": "<p>This is my first real competition. Definitely learnt quite a bit! Thanks to the organizers. After giving up on training my model, I leveraged the baseline model with few additional measures. My approach did quite well in my test cases with 2019 competition data (~0.6) but unfortunately didn't do quite as well on this contest.</p>\n<ul>\n<li><p>Precomputed embeddings for complete train data and stored in sqlite3 database. This substantially reduced time to evaluate.</p></li>\n<li><p>Mix in precomputed embeddings with embeddings generated on a subset openimages dataset. I assign a landmark ID of 999999  to this dataset and use it to filter out any images that matched this dataset. </p></li>\n<li><p>After this step, the dataset is reduced by 10x allowing a lot of time to do local matching etc. </p></li>\n</ul>\n<p>I did try using cv2 sift for local scoring but didn't see much benefit. </p>",
      "rawMarkdown": "This is my first real competition. Definitely learnt quite a bit! Thanks to the organizers. After giving up on training my model, I leveraged the baseline model with few additional measures. My approach did quite well in my test cases with 2019 competition data (~0.6) but unfortunately didn't do quite as well on this contest.\n\n* Precomputed embeddings for complete train data and stored in sqlite3 database. This substantially reduced time to evaluate.\n\n* Mix in precomputed embeddings with embeddings generated on a subset openimages dataset. I assign a landmark ID of 999999  to this dataset and use it to filter out any images that matched this dataset. \n\n* After this step, the dataset is reduced by 10x allowing a lot of time to do local matching etc. \n\nI did try using cv2 sift for local scoring but didn't see much benefit. \n",
      "votes": 5
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1032186": "This is my first real competition. Definitely learnt quite a bit! Thanks to the organizers. After giving up on training my model, I leveraged the baseline model with few additional measures. My approach did quite well in my test cases with 2019 competition data (~0.6) but unfortunately didn't do quite as well on this contest.\n\n* Precomputed embeddings for complete train data and stored in sqlite3 database. This substantially reduced time to evaluate.\n\n* Mix in precomputed embeddings with embeddings generated on a subset openimages dataset. I assign a landmark ID of 999999  to this dataset and use it to filter out any images that matched this dataset. \n\n* After this step, the dataset is reduced by 10x allowing a lot of time to do local matching etc. \n\nI did try using cv2 sift for local scoring but didn't see much benefit. \n"
  }
}