{
  "id": 53692,
  "title": "Different data augmentations using imgaug repo within Mask_RCNN",
  "url": "/competitions/data-science-bowl-2018/discussion/53692",
  "author_name": "",
  "post_date": "2018-04-03T16:53:01.394052300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi All! \nI am looking at different options for image augmentations in the imgaug repo:\n<a href=\"http://imgaug.readthedocs.io/en/latest/source/augmenters.html\">http://imgaug.readthedocs.io/en/latest/source/augmenters.html</a>\nI was wondering if someone had experience with the different options. Currently the \"OneOf \" option looks like a good choice to me: \"Apply one of a set of augmenters to each image:\"\n<a href=\"http://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof\">http://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof</a></p>\n\n<p>augmentation = imgaug.augmenters.OneOf([\n               imgaug.augmenters.Fliplr(1.0),\n               imgaug.augmenters.Affine(rotate=(-10,10))\n               imgaug.augmenters.Flipud(1.0),\n])</p>\n\n<p>Basically i am using the latest Matterport Mask_RCNN repo and adding these data augmentations. So far, Fliplr helps but the Rotation has not helped improve my results</p>\n\n<p>Thanks\nMehul</p>",
  "messages": [
    {
      "id": "308533",
      "postDate": "04/03/2018 16:53:01",
      "content": "<p>Hi All! \nI am looking at different options for image augmentations in the imgaug repo:\n<a href=\"http://imgaug.readthedocs.io/en/latest/source/augmenters.html\">http://imgaug.readthedocs.io/en/latest/source/augmenters.html</a>\nI was wondering if someone had experience with the different options. Currently the \"OneOf \" option looks like a good choice to me: \"Apply one of a set of augmenters to each image:\"\n<a href=\"http://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof\">http://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof</a></p>\n\n<p>augmentation = imgaug.augmenters.OneOf([\n               imgaug.augmenters.Fliplr(1.0),\n               imgaug.augmenters.Affine(rotate=(-10,10))\n               imgaug.augmenters.Flipud(1.0),\n])</p>\n\n<p>Basically i am using the latest Matterport Mask_RCNN repo and adding these data augmentations. So far, Fliplr helps but the Rotation has not helped improve my results</p>\n\n<p>Thanks\nMehul</p>",
      "rawMarkdown": "Hi All! \nI am looking at different options for image augmentations in the imgaug repo:\nhttp://imgaug.readthedocs.io/en/latest/source/augmenters.html\nI was wondering if someone had experience with the different options. Currently the \"OneOf \" option looks like a good choice to me: \"Apply one of a set of augmenters to each image:\"\nhttp://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof\n\naugmentation = imgaug.augmenters.OneOf([\n               imgaug.augmenters.Fliplr(1.0),\n               imgaug.augmenters.Affine(rotate=(-10,10))\n               imgaug.augmenters.Flipud(1.0),\n])\n\n\nBasically i am using the latest Matterport Mask_RCNN repo and adding these data augmentations. So far, Fliplr helps but the Rotation has not helped improve my results\n\nThanks\nMehul",
      "votes": null
    },
    {
      "id": "308824",
      "postDate": "04/04/2018 07:00:33",
      "content": "<p>Hi, \nI'm using imgaug too, but in the version pytorch of mask rcnn <a href=\"https://github.com/multimodallearning/pytorch-mask-rcnn\">https://github.com/multimodallearning/pytorch-mask-rcnn</a>  I've modified. I found your same problem. Each time I introduced the rotation, the results are worse.</p>",
      "rawMarkdown": "Hi, \nI'm using imgaug too, but in the version pytorch of mask rcnn https://github.com/multimodallearning/pytorch-mask-rcnn  I've modified. I found your same problem. Each time I introduced the rotation, the results are worse.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 308824,
      "author_name": "elenacuoco",
      "author_url": "",
      "post_date": "04/04/2018 07:00:33",
      "content": "<p>Hi, \nI'm using imgaug too, but in the version pytorch of mask rcnn <a href=\"https://github.com/multimodallearning/pytorch-mask-rcnn\">https://github.com/multimodallearning/pytorch-mask-rcnn</a>  I've modified. I found your same problem. Each time I introduced the rotation, the results are worse.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "308533": "Hi All! \nI am looking at different options for image augmentations in the imgaug repo:\nhttp://imgaug.readthedocs.io/en/latest/source/augmenters.html\nI was wondering if someone had experience with the different options. Currently the \"OneOf \" option looks like a good choice to me: \"Apply one of a set of augmenters to each image:\"\nhttp://imgaug.readthedocs.io/en/latest/source/augmenters.html#oneof\n\naugmentation = imgaug.augmenters.OneOf([\n               imgaug.augmenters.Fliplr(1.0),\n               imgaug.augmenters.Affine(rotate=(-10,10))\n               imgaug.augmenters.Flipud(1.0),\n])\n\n\nBasically i am using the latest Matterport Mask_RCNN repo and adding these data augmentations. So far, Fliplr helps but the Rotation has not helped improve my results\n\nThanks\nMehul",
    "308824": "Hi, \nI'm using imgaug too, but in the version pytorch of mask rcnn https://github.com/multimodallearning/pytorch-mask-rcnn  I've modified. I found your same problem. Each time I introduced the rotation, the results are worse."
  },
  "source": "meta"
}