{
  "id": 161858,
  "title": "Training with background noise as an additional class?",
  "url": "/competitions/birdsong-recognition/discussion/161858",
  "author_name": "",
  "post_date": "2020-06-26T12:38:17.328619Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi, just a quick question. Do you think it is worth to include some samples of just some background noise in your training data set and assign an additional class \"nocall\" to it? </p>",
  "messages": [
    {
      "id": "902904",
      "postDate": "06/26/2020 12:38:17",
      "content": "<p>Hi, just a quick question. Do you think it is worth to include some samples of just some background noise in your training data set and assign an additional class \"nocall\" to it? </p>",
      "rawMarkdown": "Hi, just a quick question. Do you think it is worth to include some samples of just some background noise in your training data set and assign an additional class \"nocall\" to it?",
      "votes": null
    },
    {
      "id": "906188",
      "postDate": "06/29/2020 05:44:41",
      "content": "<p>No that will not make sense, since normal needs to be clear and adding noise will confuse the model. X-rays are clear generally.</p>\n\n<p>What could be done is that boundary of skeleton should be considered as limit and other as outliers</p>",
      "rawMarkdown": "No that will not make sense, since normal needs to be clear and adding noise will confuse the model. X-rays are clear generally.\n\nWhat could be done is that boundary of skeleton should be considered as limit and other as outliers",
      "votes": null
    },
    {
      "id": "907970",
      "postDate": "06/30/2020 09:40:40",
      "content": "<p>I think it is a great idea, especially if we can do mix-up augmentation with some background noise clips, that can make the model more robust :)</p>",
      "rawMarkdown": "I think it is a great idea, especially if we can do mix-up augmentation with some background noise clips, that can make the model more robust :)",
      "votes": null
    },
    {
      "id": "911102",
      "postDate": "07/01/2020 15:09:17",
      "content": "<p>Finding a way of doing that effectively was part of the secret sauce for many submissions we saw for previous editions of BirdCLEF and I think mixup training is a very effective way to simulate different noise levels and the presence of anthropogenic sounds. The training data should have a lot of non-event sounds (if you find a way to extract them).</p>",
      "rawMarkdown": "Finding a way of doing that effectively was part of the secret sauce for many submissions we saw for previous editions of BirdCLEF and I think mixup training is a very effective way to simulate different noise levels and the presence of anthropogenic sounds. The training data should have a lot of non-event sounds (if you find a way to extract them).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 906188,
      "author_name": "momotabanerjee",
      "author_url": "",
      "post_date": "06/29/2020 05:44:41",
      "content": "<p>No that will not make sense, since normal needs to be clear and adding noise will confuse the model. X-rays are clear generally.</p>\n\n<p>What could be done is that boundary of skeleton should be considered as limit and other as outliers</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 907970,
      "author_name": "alanchn31",
      "author_url": "",
      "post_date": "06/30/2020 09:40:40",
      "content": "<p>I think it is a great idea, especially if we can do mix-up augmentation with some background noise clips, that can make the model more robust :)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 911102,
      "author_name": "stefankahl",
      "author_url": "",
      "post_date": "07/01/2020 15:09:17",
      "content": "<p>Finding a way of doing that effectively was part of the secret sauce for many submissions we saw for previous editions of BirdCLEF and I think mixup training is a very effective way to simulate different noise levels and the presence of anthropogenic sounds. The training data should have a lot of non-event sounds (if you find a way to extract them).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "902904": "Hi, just a quick question. Do you think it is worth to include some samples of just some background noise in your training data set and assign an additional class \"nocall\" to it?",
    "906188": "No that will not make sense, since normal needs to be clear and adding noise will confuse the model. X-rays are clear generally.\n\nWhat could be done is that boundary of skeleton should be considered as limit and other as outliers",
    "907970": "I think it is a great idea, especially if we can do mix-up augmentation with some background noise clips, that can make the model more robust :)",
    "911102": "Finding a way of doing that effectively was part of the secret sauce for many submissions we saw for previous editions of BirdCLEF and I think mixup training is a very effective way to simulate different noise levels and the presence of anthropogenic sounds. The training data should have a lot of non-event sounds (if you find a way to extract them)."
  },
  "source": "meta"
}