{
  "id": 132684,
  "title": "Faceextraction: Best models / strategies MTCNN/S3FD/BlazeFace",
  "url": "/competitions/deepfake-detection-challenge/discussion/132684",
  "author_name": "Justin Güse",
  "post_date": "2020-02-27T09:55:52.583000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hey guys, thanks in advance for everyone sharing their insights!\nWhen looking through different kernels, I noticed a lot of BlazeFace (e.g. <a href=\"https://www.kaggle.com/humananalog/starter-blazeface-pytorch\">https://www.kaggle.com/humananalog/starter-blazeface-pytorch</a>) and MTCNN (e.g. <a href=\"https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution\">https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution</a>) being used.</p>\n\n<p>But as far as I know  S3FD should be the fastest extractor right now, at least on a GPU, right?\nE.g. <a href=\"https://github.com/yxlijun/S3FD.pytorch\">https://github.com/yxlijun/S3FD.pytorch</a></p>\n\n<p>What are your experiences? Did anyone try to implement s3fd?</p>\n\n<p>And I mean there are other suggestions as well, like yolov2 for example (<a href=\"https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn\">https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn</a>)</p>",
  "messages": [
    {
      "id": 757980,
      "postDate": "2020-02-27T09:55:52.583Z",
      "content": "<p>Hey guys, thanks in advance for everyone sharing their insights!\nWhen looking through different kernels, I noticed a lot of BlazeFace (e.g. <a href=\"https://www.kaggle.com/humananalog/starter-blazeface-pytorch\">https://www.kaggle.com/humananalog/starter-blazeface-pytorch</a>) and MTCNN (e.g. <a href=\"https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution\">https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution</a>) being used.</p>\n\n<p>But as far as I know  S3FD should be the fastest extractor right now, at least on a GPU, right?\nE.g. <a href=\"https://github.com/yxlijun/S3FD.pytorch\">https://github.com/yxlijun/S3FD.pytorch</a></p>\n\n<p>What are your experiences? Did anyone try to implement s3fd?</p>\n\n<p>And I mean there are other suggestions as well, like yolov2 for example (<a href=\"https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn\">https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn</a>)</p>",
      "rawMarkdown": "Hey guys, thanks in advance for everyone sharing their insights!\nWhen looking through different kernels, I noticed a lot of BlazeFace (e.g. https://www.kaggle.com/humananalog/starter-blazeface-pytorch) and MTCNN (e.g. https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution) being used.\n\nBut as far as I know  S3FD should be the fastest extractor right now, at least on a GPU, right?\nE.g. https://github.com/yxlijun/S3FD.pytorch\n\nWhat are your experiences? Did anyone try to implement s3fd?\n\nAnd I mean there are other suggestions as well, like yolov2 for example (https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn)",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "757980": "Hey guys, thanks in advance for everyone sharing their insights!\nWhen looking through different kernels, I noticed a lot of BlazeFace (e.g. https://www.kaggle.com/humananalog/starter-blazeface-pytorch) and MTCNN (e.g. https://www.kaggle.com/timesler/fast-mtcnn-detector-55-fps-at-full-resolution) being used.\n\nBut as far as I know  S3FD should be the fastest extractor right now, at least on a GPU, right?\nE.g. https://github.com/yxlijun/S3FD.pytorch\n\nWhat are your experiences? Did anyone try to implement s3fd?\n\nAnd I mean there are other suggestions as well, like yolov2 for example (https://www.kaggle.com/basharallabadi/yolov2-vs-faced-vs-blazeface-vs-mtcnn)"
  }
}