{
  "id": 171291,
  "title": "Why less restrictions for recognition but not retrieval?",
  "url": "/competitions/landmark-recognition-2020/discussion/171291",
  "author_name": "",
  "post_date": "2020-07-31T08:03:46.783141600Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>In the retrieval challenge, we are only allowed to submit a Tensorflow 2.0 model with strict requirements on signatures and input specs, but it seems that for this recognition challenge, we could explore possibilities with local features, post-processing techniques, RANSAC, etc. and use non-TF 2.0-based solutions. I'm curious about why the organizers made a decision like this. Is it because there are more production-driving factors in a kaggle-derived retrieval solution, whereas recognition is more research oriented?</p>",
  "messages": [
    {
      "id": "952775",
      "postDate": "07/31/2020 08:03:46",
      "content": "<p>In the retrieval challenge, we are only allowed to submit a Tensorflow 2.0 model with strict requirements on signatures and input specs, but it seems that for this recognition challenge, we could explore possibilities with local features, post-processing techniques, RANSAC, etc. and use non-TF 2.0-based solutions. I'm curious about why the organizers made a decision like this. Is it because there are more production-driving factors in a kaggle-derived retrieval solution, whereas recognition is more research oriented?</p>",
      "rawMarkdown": "In the retrieval challenge, we are only allowed to submit a Tensorflow 2.0 model with strict requirements on signatures and input specs, but it seems that for this recognition challenge, we could explore possibilities with local features, post-processing techniques, RANSAC, etc. and use non-TF 2.0-based solutions. I'm curious about why the organizers made a decision like this. Is it because there are more production-driving factors in a kaggle-derived retrieval solution, whereas recognition is more research oriented?",
      "votes": null
    },
    {
      "id": "952942",
      "postDate": "07/31/2020 10:49:27",
      "content": "<p>I think that orgs would like to do an experiment with restrictions (may be they are more interested in better features), but don't want risk both competitions. Given the participation of the retrieval competition, I would call the experiment more like failure :)</p>",
      "rawMarkdown": "I think that orgs would like to do an experiment with restrictions (may be they are more interested in better features), but don't want risk both competitions. Given the participation of the retrieval competition, I would call the experiment more like failure :)",
      "votes": null
    },
    {
      "id": "955990",
      "postDate": "08/03/2020 06:16:02",
      "content": "<p>Because Tensorflow 2.0 is owned by Google and they want a Google Product solution ?\nIs this a possibility ?</p>",
      "rawMarkdown": "Because Tensorflow 2.0 is owned by Google and they want a Google Product solution ?\nIs this a possibility ?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 952942,
      "author_name": "oldufo",
      "author_url": "",
      "post_date": "07/31/2020 10:49:27",
      "content": "<p>I think that orgs would like to do an experiment with restrictions (may be they are more interested in better features), but don't want risk both competitions. Given the participation of the retrieval competition, I would call the experiment more like failure :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 955990,
      "author_name": "alincijov",
      "author_url": "",
      "post_date": "08/03/2020 06:16:02",
      "content": "<p>Because Tensorflow 2.0 is owned by Google and they want a Google Product solution ?\nIs this a possibility ?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "952775": "In the retrieval challenge, we are only allowed to submit a Tensorflow 2.0 model with strict requirements on signatures and input specs, but it seems that for this recognition challenge, we could explore possibilities with local features, post-processing techniques, RANSAC, etc. and use non-TF 2.0-based solutions. I'm curious about why the organizers made a decision like this. Is it because there are more production-driving factors in a kaggle-derived retrieval solution, whereas recognition is more research oriented?",
    "952942": "I think that orgs would like to do an experiment with restrictions (may be they are more interested in better features), but don't want risk both competitions. Given the participation of the retrieval competition, I would call the experiment more like failure :)",
    "955990": "Because Tensorflow 2.0 is owned by Google and they want a Google Product solution ?\nIs this a possibility ?"
  },
  "source": "meta"
}