{
  "id": 96795,
  "title": "Training on resized datasets",
  "url": "/competitions/open-images-2019-object-detection/discussion/96795",
  "author_name": "",
  "post_date": "2019-06-23T07:40:02.859475700Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi, I'm new to Kaggle competitions, so sorry if this question is nonsense.</p>\n\n<p>Due to my hardware limitations, I'll have to use <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/discussion/94660\">Anish's resized datasets </a>for this competition (thank you Anish for your work). But I was wondering, if I train my object detection nets on resized images, won't the test results be much worse? Because those results have to be explicitly focused on original images, right? Also, that makes me think that network architectures should be different, so I don't get how training on a resized dataset would work...</p>\n\n<p>Thank you in advance!</p>",
  "messages": [
    {
      "id": "558937",
      "postDate": "06/23/2019 07:40:02",
      "content": "<p>Hi, I'm new to Kaggle competitions, so sorry if this question is nonsense.</p>\n\n<p>Due to my hardware limitations, I'll have to use <a href=\"https://www.kaggle.com/c/open-images-2019-object-detection/discussion/94660\">Anish's resized datasets </a>for this competition (thank you Anish for your work). But I was wondering, if I train my object detection nets on resized images, won't the test results be much worse? Because those results have to be explicitly focused on original images, right? Also, that makes me think that network architectures should be different, so I don't get how training on a resized dataset would work...</p>\n\n<p>Thank you in advance!</p>",
      "rawMarkdown": "Hi, I'm new to Kaggle competitions, so sorry if this question is nonsense.\n\nDue to my hardware limitations, I'll have to use [Anish's resized datasets ](https://www.kaggle.com/c/open-images-2019-object-detection/discussion/94660)for this competition (thank you Anish for your work). But I was wondering, if I train my object detection nets on resized images, won't the test results be much worse? Because those results have to be explicitly focused on original images, right? Also, that makes me think that network architectures should be different, so I don't get how training on a resized dataset would work...\n\nThank you in advance!",
      "votes": null
    },
    {
      "id": "566555",
      "postDate": "07/02/2019 09:57:55",
      "content": "<p>I tried this at the beginning of this competition. For example, if the original images are 1024 x 768, I resized them to 512 x 384 and got map 0.384 on public LB. Now I use the same object detection net training on original size, got map 0.460.</p>",
      "rawMarkdown": "I tried this at the beginning of this competition. For example, if the original images are 1024 x 768, I resized them to 512 x 384 and got map 0.384 on public LB. Now I use the same object detection net training on original size, got map 0.460.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 566555,
      "author_name": "gujingxiao0726",
      "author_url": "",
      "post_date": "07/02/2019 09:57:55",
      "content": "<p>I tried this at the beginning of this competition. For example, if the original images are 1024 x 768, I resized them to 512 x 384 and got map 0.384 on public LB. Now I use the same object detection net training on original size, got map 0.460.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "558937": "Hi, I'm new to Kaggle competitions, so sorry if this question is nonsense.\n\nDue to my hardware limitations, I'll have to use [Anish's resized datasets ](https://www.kaggle.com/c/open-images-2019-object-detection/discussion/94660)for this competition (thank you Anish for your work). But I was wondering, if I train my object detection nets on resized images, won't the test results be much worse? Because those results have to be explicitly focused on original images, right? Also, that makes me think that network architectures should be different, so I don't get how training on a resized dataset would work...\n\nThank you in advance!",
    "566555": "I tried this at the beginning of this competition. For example, if the original images are 1024 x 768, I resized them to 512 x 384 and got map 0.384 on public LB. Now I use the same object detection net training on original size, got map 0.460."
  },
  "source": "meta"
}