{
  "id": 103916,
  "title": "score not improving",
  "url": "/competitions/aptos2019-blindness-detection/discussion/103916",
  "author_name": "",
  "post_date": "2019-08-12T21:09:52.450605700Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>What are your thoughts on coming up with an algo to remove blood vessels and the optic disc from the images? I have tried it and I find it very hard to come up with something that works with all images. \nAlso, for some reason I am not able to improve on my score (of .25) with my Resnet model (34 layers, removed black background + convert to gray scale images (Y channel equalized) + TTA + Abhishek's thresh holding technique). What else can I improve?</p>",
  "messages": [
    {
      "id": "597878",
      "postDate": "08/12/2019 21:09:52",
      "content": "<p>What are your thoughts on coming up with an algo to remove blood vessels and the optic disc from the images? I have tried it and I find it very hard to come up with something that works with all images. \nAlso, for some reason I am not able to improve on my score (of .25) with my Resnet model (34 layers, removed black background + convert to gray scale images (Y channel equalized) + TTA + Abhishek's thresh holding technique). What else can I improve?</p>",
      "rawMarkdown": "What are your thoughts on coming up with an algo to remove blood vessels and the optic disc from the images? I have tried it and I find it very hard to come up with something that works with all images. \nAlso, for some reason I am not able to improve on my score (of .25) with my Resnet model (34 layers, removed black background + convert to gray scale images (Y channel equalized) + TTA + Abhishek's thresh holding technique). What else can I improve?",
      "votes": null
    },
    {
      "id": "598494",
      "postDate": "08/13/2019 16:37:03",
      "content": "<p>You might need those items. There's a lot you can do without changing the images. \nCheck out fast.ai library.  You'll get .5 with Resnet 34 and no real changes to defaults. </p>",
      "rawMarkdown": "You might need those items. There's a lot you can do without changing the images. \nCheck out fast.ai library.  You'll get .5 with Resnet 34 and no real changes to defaults.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 598494,
      "author_name": "chrisfs",
      "author_url": "",
      "post_date": "08/13/2019 16:37:03",
      "content": "<p>You might need those items. There's a lot you can do without changing the images. \nCheck out fast.ai library.  You'll get .5 with Resnet 34 and no real changes to defaults. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "597878": "What are your thoughts on coming up with an algo to remove blood vessels and the optic disc from the images? I have tried it and I find it very hard to come up with something that works with all images. \nAlso, for some reason I am not able to improve on my score (of .25) with my Resnet model (34 layers, removed black background + convert to gray scale images (Y channel equalized) + TTA + Abhishek's thresh holding technique). What else can I improve?",
    "598494": "You might need those items. There's a lot you can do without changing the images. \nCheck out fast.ai library.  You'll get .5 with Resnet 34 and no real changes to defaults."
  },
  "source": "meta"
}