{
  "id": 228278,
  "title": "What is the best detection score only you get",
  "url": "/competitions/vinbigdata-chest-xray-abnormalities-detection/discussion/228278",
  "author_name": "David",
  "post_date": "2021-03-24T03:48:11.071000",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>I try split 70/30 and archive CV/LB 0.44/0.195 respectively. I didn't apply 2 class filter yet. I really want to know how you guys can have some good approaches to achieve really high scores.</p>",
  "messages": [
    {
      "id": 1253310,
      "postDate": "2021-03-26T15:43:14.970Z",
      "content": "<p>resnet50 and got LB 0.151, but my image size is only 256x256 for fast inference. I am wondering how you choose your CV set since the dataset is heavily biased.</p>",
      "rawMarkdown": "resnet50 and got LB 0.151, but my image size is only 256x256 for fast inference. I am wondering how you choose your CV set since the dataset is heavily biased.",
      "votes": 1,
      "replies": [
        {
          "id": 1255602,
          "postDate": "2021-03-29T00:37:26.750Z",
          "content": "<p>I think in order to get high mAP you should choose some SOTA models instead of resnet50, and the size should be a little bit larger ex 512<em>512, 768</em>768, 1024*1024…. The CV I split 70-30 based on the distribution of every single class.</p>",
          "rawMarkdown": "I think in order to get high mAP you should choose some SOTA models instead of resnet50, and the size should be a little bit larger ex 512*512, 768*768, 1024*1024.... The CV I split 70-30 based on the distribution of every single class."
        }
      ]
    },
    {
      "id": 1251647,
      "postDate": "2021-03-25T01:28:29.337Z",
      "content": "<p>yolov5 cv 0.41 / lb 0.165 without 2 class filter<br>\nwhat model do you use?</p>",
      "rawMarkdown": "yolov5 cv 0.41 / lb 0.165 without 2 class filter\nwhat model do you use?",
      "votes": 1,
      "replies": [
        {
          "id": 1251650,
          "postDate": "2021-03-25T01:46:22.867Z",
          "content": "<p>I use COCO SOTA scaled yolov4 for this model. You can also improve the score by using model ensemble technique. You're 50th position seem your 2 class filter is so good</p>",
          "rawMarkdown": "I use COCO SOTA scaled yolov4 for this model. You can also improve the score by using model ensemble technique. You're 50th position seem your 2 class filter is so good",
          "votes": 1
        },
        {
          "id": 1251699,
          "postDate": "2021-03-25T03:32:40.597Z",
          "content": "<p>Seems like your 2 class filter increase the score up to 0.1</p>",
          "rawMarkdown": "Seems like your 2 class filter increase the score up to 0.1",
          "votes": 1
        },
        {
          "id": 1251750,
          "postDate": "2021-03-25T05:09:31.503Z",
          "content": "<p>But there is only 10% on LB so we don't have some good conclusion now. For me, I more trust on CV</p>",
          "rawMarkdown": "But there is only 10% on LB so we don't have some good conclusion now. For me, I more trust on CV"
        }
      ]
    },
    {
      "id": 1250443,
      "postDate": "2021-03-24T03:48:11.070Z",
      "content": "<p>I try split 70/30 and archive CV/LB 0.44/0.195 respectively. I didn't apply 2 class filter yet. I really want to know how you guys can have some good approaches to achieve really high scores.</p>",
      "rawMarkdown": "I try split 70/30 and archive CV/LB 0.44/0.195 respectively. I didn't apply 2 class filter yet. I really want to know how you guys can have some good approaches to achieve really high scores.",
      "votes": 2
    }
  ],
  "comments": [
    {
      "id": 1253310,
      "author_name": "kidd",
      "author_url": "",
      "post_date": "2021-03-26T15:43:14.970000",
      "content": "<p>resnet50 and got LB 0.151, but my image size is only 256x256 for fast inference. I am wondering how you choose your CV set since the dataset is heavily biased.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1255602,
          "author_name": "David",
          "author_url": "",
          "post_date": "2021-03-29T00:37:26.750000",
          "content": "<p>I think in order to get high mAP you should choose some SOTA models instead of resnet50, and the size should be a little bit larger ex 512<em>512, 768</em>768, 1024*1024…. The CV I split 70-30 based on the distribution of every single class.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1251647,
      "author_name": "kaggler",
      "author_url": "",
      "post_date": "2021-03-25T01:28:29.337000",
      "content": "<p>yolov5 cv 0.41 / lb 0.165 without 2 class filter<br>\nwhat model do you use?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1251650,
          "author_name": "David",
          "author_url": "",
          "post_date": "2021-03-25T01:46:22.867000",
          "content": "<p>I use COCO SOTA scaled yolov4 for this model. You can also improve the score by using model ensemble technique. You're 50th position seem your 2 class filter is so good</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1251699,
          "author_name": "Phat Tran",
          "author_url": "",
          "post_date": "2021-03-25T03:32:40.597000",
          "content": "<p>Seems like your 2 class filter increase the score up to 0.1</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1251750,
          "author_name": "David",
          "author_url": "",
          "post_date": "2021-03-25T05:09:31.503000",
          "content": "<p>But there is only 10% on LB so we don't have some good conclusion now. For me, I more trust on CV</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1253310": "resnet50 and got LB 0.151, but my image size is only 256x256 for fast inference. I am wondering how you choose your CV set since the dataset is heavily biased.",
    "1251647": "yolov5 cv 0.41 / lb 0.165 without 2 class filter\nwhat model do you use?",
    "1250443": "I try split 70/30 and archive CV/LB 0.44/0.195 respectively. I didn't apply 2 class filter yet. I really want to know how you guys can have some good approaches to achieve really high scores."
  }
}