{
  "id": 102332,
  "title": "Ideas for finding the matching eye?",
  "url": "/competitions/aptos2019-blindness-detection/discussion/102332",
  "author_name": "",
  "post_date": "2019-08-01T12:57:43.565683900Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Reading the solutions of the past competition, I realized that the 3 top teams combined the predictions of both eyes. I wonder if the doctor assigned the labels for the person or for each eye. What do you think?\nSince in this competition we don't know which is the right and left eye for each person, it would probably be a boost for the model to identify the most probable matching eye to blend the predictions.\nAny ideas about the best way to do this?</p>",
  "messages": [
    {
      "id": "589839",
      "postDate": "08/01/2019 12:57:43",
      "content": "<p>Reading the solutions of the past competition, I realized that the 3 top teams combined the predictions of both eyes. I wonder if the doctor assigned the labels for the person or for each eye. What do you think?\nSince in this competition we don't know which is the right and left eye for each person, it would probably be a boost for the model to identify the most probable matching eye to blend the predictions.\nAny ideas about the best way to do this?</p>",
      "rawMarkdown": "Reading the solutions of the past competition, I realized that the 3 top teams combined the predictions of both eyes. I wonder if the doctor assigned the labels for the person or for each eye. What do you think?\nSince in this competition we don't know which is the right and left eye for each person, it would probably be a boost for the model to identify the most probable matching eye to blend the predictions.\nAny ideas about the best way to do this?",
      "votes": null
    },
    {
      "id": "590045",
      "postDate": "08/01/2019 18:29:54",
      "content": "<p>May be you can train a model on previous competition data, so that it can detect left and right eye of same person and use it to pair images of this competition. Anyways retinas have quite unique signature. Something like triplet loss or similar can do the job. If you find something useful, please do share (if you wish 😊 ). All the best.</p>",
      "rawMarkdown": "May be you can train a model on previous competition data, so that it can detect left and right eye of same person and use it to pair images of this competition. Anyways retinas have quite unique signature. Something like triplet loss or similar can do the job. If you find something useful, please do share (if you wish 😊 ). All the best.",
      "votes": null
    },
    {
      "id": "590176",
      "postDate": "08/01/2019 21:16:14",
      "content": "<p>Yes, that seems to be the best approach. Maybe averaging the prediction of all the samples with the highest probabilities could work. Another idea is averaging the nearest neigbors, but in this case it will probably consider the same side eye.</p>",
      "rawMarkdown": "Yes, that seems to be the best approach. Maybe averaging the prediction of all the samples with the highest probabilities could work. Another idea is averaging the nearest neigbors, but in this case it will probably consider the same side eye.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 590045,
      "author_name": "ashwan1",
      "author_url": "",
      "post_date": "08/01/2019 18:29:54",
      "content": "<p>May be you can train a model on previous competition data, so that it can detect left and right eye of same person and use it to pair images of this competition. Anyways retinas have quite unique signature. Something like triplet loss or similar can do the job. If you find something useful, please do share (if you wish 😊 ). All the best.</p>",
      "votes": null,
      "replies": [
        {
          "id": 590176,
          "author_name": "carlospk",
          "author_url": "",
          "post_date": "08/01/2019 21:16:14",
          "content": "<p>Yes, that seems to be the best approach. Maybe averaging the prediction of all the samples with the highest probabilities could work. Another idea is averaging the nearest neigbors, but in this case it will probably consider the same side eye.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "589839": "Reading the solutions of the past competition, I realized that the 3 top teams combined the predictions of both eyes. I wonder if the doctor assigned the labels for the person or for each eye. What do you think?\nSince in this competition we don't know which is the right and left eye for each person, it would probably be a boost for the model to identify the most probable matching eye to blend the predictions.\nAny ideas about the best way to do this?",
    "590045": "May be you can train a model on previous competition data, so that it can detect left and right eye of same person and use it to pair images of this competition. Anyways retinas have quite unique signature. Something like triplet loss or similar can do the job. If you find something useful, please do share (if you wish 😊 ). All the best.",
    "590176": "Yes, that seems to be the best approach. Maybe averaging the prediction of all the samples with the highest probabilities could work. Another idea is averaging the nearest neigbors, but in this case it will probably consider the same side eye."
  },
  "source": "meta"
}