{
  "id": 20280,
  "title": "Repetitive training images - good or bad?",
  "url": "/competitions/state-farm-distracted-driver-detection/discussion/20280",
  "author_name": "",
  "post_date": "2016-04-20T12:08:48.033Z",
  "votes": 1,
  "comment_count": 1,
  "views": 539,
  "content": "<p>There are very few subjects whose images are present in the training set. There are an average of around 2100 images in each training classes. I was browsing through them and found that there are multiple images in a given class of the same person. So, if we talk about say class c6, &quot;Bob&quot; may be drinking coffee in about 20 of the images in c6. </p>\n\n<p>Would this cause some unwanted bias in the model getting trained? Would removing such repetitive images yield better results? I'd like the opinion of other participants.</p>",
  "messages": [
    {
      "id": "115833",
      "postDate": "04/20/2016 12:08:48",
      "content": "<p>There are very few subjects whose images are present in the training set. There are an average of around 2100 images in each training classes. I was browsing through them and found that there are multiple images in a given class of the same person. So, if we talk about say class c6, &quot;Bob&quot; may be drinking coffee in about 20 of the images in c6. </p>\n\n<p>Would this cause some unwanted bias in the model getting trained? Would removing such repetitive images yield better results? I'd like the opinion of other participants.</p>",
      "rawMarkdown": "There are very few subjects whose images are present in the training set. There are an average of around 2100 images in each training classes. I was browsing through them and found that there are multiple images in a given class of the same person. So, if we talk about say class c6, \"Bob\" may be drinking coffee in about 20 of the images in c6. \r\n\r\nWould this cause some unwanted bias in the model getting trained? Would removing such repetitive images yield better results? I'd like the opinion of other participants.",
      "votes": null
    },
    {
      "id": "116855",
      "postDate": "04/26/2016 10:02:50",
      "content": "<p>You should not remove the repetitions because even though it is the same person and the same action, the images are different (human pose can not be exacltly the same), it helps to increase variability in the data.</p>",
      "rawMarkdown": "You should not remove the repetitions because even though it is the same person and the same action, the images are different (human pose can not be exacltly the same), it helps to increase variability in the data.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 116855,
      "author_name": "biagio90",
      "author_url": "",
      "post_date": "04/26/2016 10:02:50",
      "content": "<p>You should not remove the repetitions because even though it is the same person and the same action, the images are different (human pose can not be exacltly the same), it helps to increase variability in the data.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "115833": "There are very few subjects whose images are present in the training set. There are an average of around 2100 images in each training classes. I was browsing through them and found that there are multiple images in a given class of the same person. So, if we talk about say class c6, \"Bob\" may be drinking coffee in about 20 of the images in c6. \r\n\r\nWould this cause some unwanted bias in the model getting trained? Would removing such repetitive images yield better results? I'd like the opinion of other participants.",
    "116855": "You should not remove the repetitions because even though it is the same person and the same action, the images are different (human pose can not be exacltly the same), it helps to increase variability in the data."
  },
  "source": "meta"
}