{
  "id": 169311,
  "title": "Any idea about data clean",
  "url": "/competitions/global-wheat-detection/discussion/169311",
  "author_name": "",
  "post_date": "2020-07-23T14:08:59.529176200Z",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<ul>\n<li>I have created a <a href=\"https://www.kaggle.com/raininbox/check-clean-big-small-bboxes/comments\">notebook</a> for filter the boxes with abnormal size (w/h&lt;10 or area&lt;300 or huge box) </li>\n<li>Any other clean data idea? such as : confident learning(<a href=\"https://github.com/cgnorthcutt/cleanlab\">github</a>)</li>\n</ul>",
  "messages": [
    {
      "id": "941955",
      "postDate": "07/23/2020 14:08:59",
      "content": "<ul>\n<li>I have created a <a href=\"https://www.kaggle.com/raininbox/check-clean-big-small-bboxes/comments\">notebook</a> for filter the boxes with abnormal size (w/h&lt;10 or area&lt;300 or huge box) </li>\n<li>Any other clean data idea? such as : confident learning(<a href=\"https://github.com/cgnorthcutt/cleanlab\">github</a>)</li>\n</ul>",
      "rawMarkdown": "I have created a [notebook](https://www.kaggle.com/raininbox/check-clean-big-small-bboxes/comments) for filter the boxes with abnormal size (w/h&lt;10 or area&lt;300 or huge box) \n- Any other clean data idea? such as : confident learning([github](https://github.com/cgnorthcutt/cleanlab))",
      "votes": null
    },
    {
      "id": "942877",
      "postDate": "07/24/2020 03:41:51",
      "content": "<p>I filtered boxes which had an extreme height/width ratio too. Also, since I applied augmentations, I had two additional criteria:\n1) the mean color of each box is not too bright or dim\n2) the box area after cropping is atleast 10% of the original box area</p>",
      "rawMarkdown": "I filtered boxes which had an extreme height/width ratio too. Also, since I applied augmentations, I had two additional criteria:\n1) the mean color of each box is not too bright or dim\n2) the box area after cropping is atleast 10% of the original box area",
      "votes": null
    },
    {
      "id": "943043",
      "postDate": "07/24/2020 05:41:32",
      "content": "<p>That must be a point to boost.What's more, the pseduo label will make the model fight agaisnt the error lables.make the epoch longer :)</p>",
      "rawMarkdown": "That must be a point to boost.What's more, the pseduo label will make the model fight agaisnt the error lables.make the epoch longer :)",
      "votes": null
    },
    {
      "id": "943311",
      "postDate": "07/24/2020 09:28:21",
      "content": "<p><a href=\"/raininbox\">@raininbox</a> \ndid u see any score drop after using cleaned labels ?</p>",
      "rawMarkdown": "raininbox \ndid u see any score drop after using cleaned labels ?",
      "votes": null
    },
    {
      "id": "943493",
      "postDate": "07/24/2020 11:59:30",
      "content": "<p>😄 It needs train a CNN model to classify the background and postground(wheat). I haven't tried it yet</p>",
      "rawMarkdown": "😄 It needs train a CNN model to classify the background and postground(wheat). I haven't tried it yet",
      "votes": null
    },
    {
      "id": "943774",
      "postDate": "07/24/2020 15:11:34",
      "content": "<p><a href=\"/raininbox\">@raininbox</a> </p>\n\n<p>I have done the same thing as you: filter large/small boxes and uncommon size ratio boxes.</p>\n\n<p>This paper looks quite interesting and I wish I have the time to try ideas in the paper. 😩 \n<a href=\"https://arxiv.org/abs/2003.01285\">Towards Noise-resistant Object Detection with Noisy Annotations</a></p>",
      "rawMarkdown": "raininbox \n\nI have done the same thing as you: filter large/small boxes and uncommon size ratio boxes.\n\nThis paper looks quite interesting and I wish I have the time to try ideas in the paper. 😩 \n[Towards Noise-resistant Object Detection with Noisy Annotations](https://arxiv.org/abs/2003.01285)",
      "votes": null
    },
    {
      "id": "943776",
      "postDate": "07/24/2020 15:12:14",
      "content": "<p>I have trained with clean box labels and haven't seen any significant performance drop.\nPS: My training pipeline is based on shonenkov's efficientdet <a href=\"https://www.kaggle.com/shonenkov/training-efficientdet\">notebook</a></p>",
      "rawMarkdown": "I have trained with clean box labels and haven't seen any significant performance drop.\nPS: My training pipeline is based on shonenkov's efficientdet [notebook](https://www.kaggle.com/shonenkov/training-efficientdet)",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 942877,
      "author_name": "matthewmasters",
      "author_url": "",
      "post_date": "07/24/2020 03:41:51",
      "content": "<p>I filtered boxes which had an extreme height/width ratio too. Also, since I applied augmentations, I had two additional criteria:\n1) the mean color of each box is not too bright or dim\n2) the box area after cropping is atleast 10% of the original box area</p>",
      "votes": null,
      "replies": [
        {
          "id": 943043,
          "author_name": "raininbox",
          "author_url": "",
          "post_date": "07/24/2020 05:41:32",
          "content": "<p>That must be a point to boost.What's more, the pseduo label will make the model fight agaisnt the error lables.make the epoch longer :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 943311,
      "author_name": "jaideepvalani",
      "author_url": "",
      "post_date": "07/24/2020 09:28:21",
      "content": "<p><a href=\"/raininbox\">@raininbox</a> \ndid u see any score drop after using cleaned labels ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 943493,
          "author_name": "raininbox",
          "author_url": "",
          "post_date": "07/24/2020 11:59:30",
          "content": "<p>😄 It needs train a CNN model to classify the background and postground(wheat). I haven't tried it yet</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 943776,
          "author_name": "kuohsintu",
          "author_url": "",
          "post_date": "07/24/2020 15:12:14",
          "content": "<p>I have trained with clean box labels and haven't seen any significant performance drop.\nPS: My training pipeline is based on shonenkov's efficientdet <a href=\"https://www.kaggle.com/shonenkov/training-efficientdet\">notebook</a></p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 943774,
      "author_name": "kuohsintu",
      "author_url": "",
      "post_date": "07/24/2020 15:11:34",
      "content": "<p><a href=\"/raininbox\">@raininbox</a> </p>\n\n<p>I have done the same thing as you: filter large/small boxes and uncommon size ratio boxes.</p>\n\n<p>This paper looks quite interesting and I wish I have the time to try ideas in the paper. 😩 \n<a href=\"https://arxiv.org/abs/2003.01285\">Towards Noise-resistant Object Detection with Noisy Annotations</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "941955": "I have created a [notebook](https://www.kaggle.com/raininbox/check-clean-big-small-bboxes/comments) for filter the boxes with abnormal size (w/h&lt;10 or area&lt;300 or huge box) \n- Any other clean data idea? such as : confident learning([github](https://github.com/cgnorthcutt/cleanlab))",
    "942877": "I filtered boxes which had an extreme height/width ratio too. Also, since I applied augmentations, I had two additional criteria:\n1) the mean color of each box is not too bright or dim\n2) the box area after cropping is atleast 10% of the original box area",
    "943043": "That must be a point to boost.What's more, the pseduo label will make the model fight agaisnt the error lables.make the epoch longer :)",
    "943311": "raininbox \ndid u see any score drop after using cleaned labels ?",
    "943493": "😄 It needs train a CNN model to classify the background and postground(wheat). I haven't tried it yet",
    "943774": "raininbox \n\nI have done the same thing as you: filter large/small boxes and uncommon size ratio boxes.\n\nThis paper looks quite interesting and I wish I have the time to try ideas in the paper. 😩 \n[Towards Noise-resistant Object Detection with Noisy Annotations](https://arxiv.org/abs/2003.01285)",
    "943776": "I have trained with clean box labels and haven't seen any significant performance drop.\nPS: My training pipeline is based on shonenkov's efficientdet [notebook](https://www.kaggle.com/shonenkov/training-efficientdet)"
  },
  "source": "meta"
}