{
  "id": 80077,
  "title": "can we use face recognition approach?",
  "url": "/competitions/humpback-whale-identification/discussion/80077",
  "author_name": "",
  "post_date": "2019-02-10T10:23:16.178524100Z",
  "votes": 8,
  "comment_count": 6,
  "views": 0,
  "content": "<p>some ideas from old face recognition papers:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11236/Slide12.png\" alt=\"enter image description here\"></p>\n\n<p>this make feature multi scale and robust to occlusion </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11235/Slide14.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11237/Slide13.png\" alt=\"enter image description here\"></p>\n\n<p>our problem is \"close\" open set face identification problem.</p>",
  "messages": [
    {
      "id": "469020",
      "postDate": "02/10/2019 10:23:16",
      "content": "<p>some ideas from old face recognition papers:</p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11236/Slide12.png\" alt=\"enter image description here\"></p>\n\n<p>this make feature multi scale and robust to occlusion </p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11235/Slide14.png\" alt=\"enter image description here\"></p>\n\n<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11237/Slide13.png\" alt=\"enter image description here\"></p>\n\n<p>our problem is \"close\" open set face identification problem.</p>",
      "rawMarkdown": "some ideas from old face recognition papers:\n\n  ![enter image description here][1]\n\n this make feature multi scale and robust to occlusion \n\n  ![enter image description here][2]\n\n  ![enter image description here][3]\n\nour problem is \"close\" open set face identification problem.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11236/Slide12.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11235/Slide14.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11237/Slide13.png",
      "votes": null
    },
    {
      "id": "469031",
      "postDate": "02/10/2019 10:56:50",
      "content": "<p>triple loss is difficult to train. I suggest replace it with large margin cluster/center loss, e.g. look at :\n<a href=\"https://github.com/YirongMao/softmax_variants\">https://github.com/YirongMao/softmax_variants</a></p>",
      "rawMarkdown": "triple loss is difficult to train. I suggest replace it with large margin cluster/center loss, e.g. look at :\nhttps://github.com/YirongMao/softmax_variants",
      "votes": null
    },
    {
      "id": "469096",
      "postDate": "02/10/2019 13:41:22",
      "content": "<p>wow very interesting approach.</p>",
      "rawMarkdown": "wow very interesting approach.",
      "votes": null
    },
    {
      "id": "469346",
      "postDate": "02/11/2019 02:57:27",
      "content": "<p>My experimentation with center loss hasn't been successful. It seems that given the small amount of data available in this competition, center loss would significantly overfit the training data by clustering the few train images for each whale. I might need to tune some parameters to make it less aggressive.</p>",
      "rawMarkdown": "My experimentation with center loss hasn't been successful. It seems that given the small amount of data available in this competition, center loss would significantly overfit the training data by clustering the few train images for each whale. I might need to tune some parameters to make it less aggressive.",
      "votes": null
    },
    {
      "id": "469352",
      "postDate": "02/11/2019 03:05:16",
      "content": "<p>use  heavy augmentation.</p>\n\n<p>the input varies mostly by affine transform and i think it could be well modeled. Affine transform are  image manifold.</p>\n\n<p>you can also try large margin gaussian mixture loss, which is an improvement over center loss. Try map the feature of each class into standard gaussian distribution (covariance =1)</p>",
      "rawMarkdown": "use  heavy augmentation.\n\nthe input varies mostly by affine transform and i think it could be well modeled. Affine transform are  image manifold.\n\nyou can also try large margin gaussian mixture loss, which is an improvement over center loss. Try map the feature of each class into standard gaussian distribution (covariance =1)",
      "votes": null
    },
    {
      "id": "469427",
      "postDate": "02/11/2019 07:24:01",
      "content": "<p>Here I find a website that illustrate Modern Face Recognition with Deep Learning\n<a href=\"https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78\">https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78</a>\nAccording to this, if we can find the keypoints of whale tails during the pre-processing stage, things will be much more easier.</p>",
      "rawMarkdown": "Here I find a website that illustrate Modern Face Recognition with Deep Learning\nhttps://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78\nAccording to this, if we can find the keypoints of whale tails during the pre-processing stage, things will be much more easier.",
      "votes": null
    },
    {
      "id": "469455",
      "postDate": "02/11/2019 08:42:25",
      "content": "<p>yes, that is the trick in face recognition. some paper report experiment results with and without alignment.</p>\n\n<p>if the image is well aligned, metric learning is simple because the distribution per clause become more Gaussian. Affine transform or pose change is a complex manifold</p>",
      "rawMarkdown": "yes, that is the trick in face recognition. some paper report experiment results with and without alignment.\n\nif the image is well aligned, metric learning is simple because the distribution per clause become more Gaussian. Affine transform or pose change is a complex manifold",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 469031,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "02/10/2019 10:56:50",
      "content": "<p>triple loss is difficult to train. I suggest replace it with large margin cluster/center loss, e.g. look at :\n<a href=\"https://github.com/YirongMao/softmax_variants\">https://github.com/YirongMao/softmax_variants</a></p>",
      "votes": null,
      "replies": [
        {
          "id": 469346,
          "author_name": "mingzhao03",
          "author_url": "",
          "post_date": "02/11/2019 02:57:27",
          "content": "<p>My experimentation with center loss hasn't been successful. It seems that given the small amount of data available in this competition, center loss would significantly overfit the training data by clustering the few train images for each whale. I might need to tune some parameters to make it less aggressive.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 469352,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "02/11/2019 03:05:16",
          "content": "<p>use  heavy augmentation.</p>\n\n<p>the input varies mostly by affine transform and i think it could be well modeled. Affine transform are  image manifold.</p>\n\n<p>you can also try large margin gaussian mixture loss, which is an improvement over center loss. Try map the feature of each class into standard gaussian distribution (covariance =1)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 469096,
      "author_name": "dhaqui",
      "author_url": "",
      "post_date": "02/10/2019 13:41:22",
      "content": "<p>wow very interesting approach.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 469427,
      "author_name": "dilapsky",
      "author_url": "",
      "post_date": "02/11/2019 07:24:01",
      "content": "<p>Here I find a website that illustrate Modern Face Recognition with Deep Learning\n<a href=\"https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78\">https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78</a>\nAccording to this, if we can find the keypoints of whale tails during the pre-processing stage, things will be much more easier.</p>",
      "votes": null,
      "replies": [
        {
          "id": 469455,
          "author_name": "hengck23",
          "author_url": "",
          "post_date": "02/11/2019 08:42:25",
          "content": "<p>yes, that is the trick in face recognition. some paper report experiment results with and without alignment.</p>\n\n<p>if the image is well aligned, metric learning is simple because the distribution per clause become more Gaussian. Affine transform or pose change is a complex manifold</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "469020": "some ideas from old face recognition papers:\n\n  ![enter image description here][1]\n\n this make feature multi scale and robust to occlusion \n\n  ![enter image description here][2]\n\n  ![enter image description here][3]\n\nour problem is \"close\" open set face identification problem.\n\n\n  [1]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11236/Slide12.png\n  [2]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11235/Slide14.png\n  [3]: https://storage.googleapis.com/kaggle-forum-message-attachments/469020/11237/Slide13.png",
    "469031": "triple loss is difficult to train. I suggest replace it with large margin cluster/center loss, e.g. look at :\nhttps://github.com/YirongMao/softmax_variants",
    "469096": "wow very interesting approach.",
    "469346": "My experimentation with center loss hasn't been successful. It seems that given the small amount of data available in this competition, center loss would significantly overfit the training data by clustering the few train images for each whale. I might need to tune some parameters to make it less aggressive.",
    "469352": "use  heavy augmentation.\n\nthe input varies mostly by affine transform and i think it could be well modeled. Affine transform are  image manifold.\n\nyou can also try large margin gaussian mixture loss, which is an improvement over center loss. Try map the feature of each class into standard gaussian distribution (covariance =1)",
    "469427": "Here I find a website that illustrate Modern Face Recognition with Deep Learning\nhttps://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78\nAccording to this, if we can find the keypoints of whale tails during the pre-processing stage, things will be much more easier.",
    "469455": "yes, that is the trick in face recognition. some paper report experiment results with and without alignment.\n\nif the image is well aligned, metric learning is simple because the distribution per clause become more Gaussian. Affine transform or pose change is a complex manifold"
  },
  "source": "meta"
}