{
  "id": 55602,
  "title": "Any results with a discriminative loss function?",
  "url": "/competitions/data-science-bowl-2018/discussion/55602",
  "author_name": "",
  "post_date": "2018-04-29T11:37:47.946652900Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Did anyone get good results with a discriminative loss function? As described in this article\n<a href=\"https://arxiv.org/abs/1708.02551\">https://arxiv.org/abs/1708.02551</a></p>\n\n<p>I implemented this solution during the competition, however, the performance was significantly worse than both my MaskRCNN and UNet based solution. I have tried to work on this solution more, but it is still training very slowly.</p>\n\n<p>I am wondering if it is my implementation and someone else got good results using this.</p>\n\n<p>I found this article, where the DSB18 data is used, but it does not go into detail about the performance.\n<a href=\"https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1\">https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1</a></p>",
  "messages": [
    {
      "id": "320642",
      "postDate": "04/29/2018 11:37:47",
      "content": "<p>Did anyone get good results with a discriminative loss function? As described in this article\n<a href=\"https://arxiv.org/abs/1708.02551\">https://arxiv.org/abs/1708.02551</a></p>\n\n<p>I implemented this solution during the competition, however, the performance was significantly worse than both my MaskRCNN and UNet based solution. I have tried to work on this solution more, but it is still training very slowly.</p>\n\n<p>I am wondering if it is my implementation and someone else got good results using this.</p>\n\n<p>I found this article, where the DSB18 data is used, but it does not go into detail about the performance.\n<a href=\"https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1\">https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1</a></p>",
      "rawMarkdown": "Did anyone get good results with a discriminative loss function? As described in this article\n[https://arxiv.org/abs/1708.02551][1]\n\nI implemented this solution during the competition, however, the performance was significantly worse than both my MaskRCNN and UNet based solution. I have tried to work on this solution more, but it is still training very slowly.\n\nI am wondering if it is my implementation and someone else got good results using this.\n\nI found this article, where the DSB18 data is used, but it does not go into detail about the performance.\n[https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1][2]\n\n\n  [1]: https://arxiv.org/abs/1708.02551\n  [2]: https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1",
      "votes": null
    },
    {
      "id": "502970",
      "postDate": "03/29/2019 10:25:47",
      "content": "<p>I also tried to make some out of it, but at current moment there is no luck. But I've modified discriminative loss to simplifying embedding space and it works much better. The main idea - to get rid of clusters push force and regularization force, because it make things very difficult to fit. But instead just fix the cluster centers and apply pull force. </p>",
      "rawMarkdown": "I also tried to make some out of it, but at current moment there is no luck. But I've modified discriminative loss to simplifying embedding space and it works much better. The main idea - to get rid of clusters push force and regularization force, because it make things very difficult to fit. But instead just fix the cluster centers and apply pull force.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 502970,
      "author_name": "sprocketus",
      "author_url": "",
      "post_date": "03/29/2019 10:25:47",
      "content": "<p>I also tried to make some out of it, but at current moment there is no luck. But I've modified discriminative loss to simplifying embedding space and it works much better. The main idea - to get rid of clusters push force and regularization force, because it make things very difficult to fit. But instead just fix the cluster centers and apply pull force. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "320642": "Did anyone get good results with a discriminative loss function? As described in this article\n[https://arxiv.org/abs/1708.02551][1]\n\nI implemented this solution during the competition, however, the performance was significantly worse than both my MaskRCNN and UNet based solution. I have tried to work on this solution more, but it is still training very slowly.\n\nI am wondering if it is my implementation and someone else got good results using this.\n\nI found this article, where the DSB18 data is used, but it does not go into detail about the performance.\n[https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1][2]\n\n\n  [1]: https://arxiv.org/abs/1708.02551\n  [2]: https://towardsdatascience.com/instance-embedding-instance-segmentation-without-proposals-31946a7c53e1",
    "502970": "I also tried to make some out of it, but at current moment there is no luck. But I've modified discriminative loss to simplifying embedding space and it works much better. The main idea - to get rid of clusters push force and regularization force, because it make things very difficult to fit. But instead just fix the cluster centers and apply pull force."
  },
  "source": "meta"
}