{
  "id": 208367,
  "title": "TTA for Object detection",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/208367",
  "author_name": "DeepUnderstanding",
  "post_date": "2021-01-03T06:15:39.312000",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Generally in image classifications, we generally average the results coming from the different augmentations.<br>\nBut if we do some simple augmentations like Flipping to Transposing it will give different coordinate for the same object detected.<br>\nSo how do we do TTA then?<br>\nThank You.</p>",
  "messages": [
    {
      "id": 1138171,
      "postDate": "2021-01-04T13:17:33.363Z",
      "content": "<p><a href=\"https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection\" target=\"_blank\">https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection</a></p>\n<p>Hope it helps</p>",
      "rawMarkdown": "https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection\n\nHope it helps",
      "votes": 2
    },
    {
      "id": 1136926,
      "postDate": "2021-01-03T14:39:35.937Z",
      "content": "<ul>\n<li>Flip the image (or use any other augmentation)</li>\n<li>Predict</li>\n<li>\"unaugment\" the predicted boxes (flip back)</li>\n<li>After the boxes are the same place/size/etc you can merge/average them (nms, wbf, or with some similar technique)</li>\n</ul>\n<p>I am working on a general implementation, I'll share it soon.</p>",
      "rawMarkdown": "- Flip the image (or use any other augmentation)\n- Predict\n- \"unaugment\" the predicted boxes (flip back)\n- After the boxes are the same place/size/etc you can merge/average them (nms, wbf, or with some similar technique)\n\nI am working on a general implementation, I'll share it soon.",
      "votes": 2
    },
    {
      "id": 1136489,
      "postDate": "2021-01-03T06:15:39.313Z",
      "content": "<p>Generally in image classifications, we generally average the results coming from the different augmentations.<br>\nBut if we do some simple augmentations like Flipping to Transposing it will give different coordinate for the same object detected.<br>\nSo how do we do TTA then?<br>\nThank You.</p>",
      "rawMarkdown": "Generally in image classifications, we generally average the results coming from the different augmentations.\nBut if we do some simple augmentations like Flipping to Transposing it will give different coordinate for the same object detected.\nSo how do we do TTA then?\nThank You.",
      "votes": 2
    },
    {
      "id": 1136869,
      "postDate": "2021-01-03T13:56:05.687Z",
      "content": "<p>add rotation , horizontal flip</p>",
      "rawMarkdown": "add rotation , horizontal flip"
    }
  ],
  "comments": [
    {
      "id": 1138171,
      "author_name": "Phat Tran",
      "author_url": "",
      "post_date": "2021-01-04T13:17:33.363000",
      "content": "<p><a href=\"https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection\" target=\"_blank\">https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection</a></p>\n<p>Hope it helps</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1136926,
      "author_name": "Peter",
      "author_url": "",
      "post_date": "2021-01-03T14:39:35.937000",
      "content": "<ul>\n<li>Flip the image (or use any other augmentation)</li>\n<li>Predict</li>\n<li>\"unaugment\" the predicted boxes (flip back)</li>\n<li>After the boxes are the same place/size/etc you can merge/average them (nms, wbf, or with some similar technique)</li>\n</ul>\n<p>I am working on a general implementation, I'll share it soon.</p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1136869,
      "author_name": "Mohammed Rizin V K",
      "author_url": "",
      "post_date": "2021-01-03T13:56:05.687000",
      "content": "<p>add rotation , horizontal flip</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1138171": "https://github.com/ancasag/ensembleObjectDetection#test-time-augmentation-for-object-detection\n\nHope it helps",
    "1136926": "- Flip the image (or use any other augmentation)\n- Predict\n- \"unaugment\" the predicted boxes (flip back)\n- After the boxes are the same place/size/etc you can merge/average them (nms, wbf, or with some similar technique)\n\nI am working on a general implementation, I'll share it soon.",
    "1136489": "Generally in image classifications, we generally average the results coming from the different augmentations.\nBut if we do some simple augmentations like Flipping to Transposing it will give different coordinate for the same object detected.\nSo how do we do TTA then?\nThank You.",
    "1136869": "add rotation , horizontal flip"
  }
}