{
  "id": 242696,
  "title": "how to use images without bounding boxes?",
  "url": "/competitions/siim-covid19-detection/discussion/242696",
  "author_name": "",
  "post_date": "2021-05-30T09:49:59.383184500Z",
  "votes": 3,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I train an object detector based on <a href=\"https://www.kaggle.com/shonenkov/timm-efficientdet-pytorch\" target=\"_blank\">efficientdet</a>,but some images don't have bounding boxes,so I don't know how to use those images correctly.If I should label those images zero,and train those images as well as other images,or throw those images directly?</p>",
  "messages": [
    {
      "id": "1328485",
      "postDate": "05/30/2021 09:49:59",
      "content": "<p>I train an object detector based on <a href=\"https://www.kaggle.com/shonenkov/timm-efficientdet-pytorch\" target=\"_blank\">efficientdet</a>,but some images don't have bounding boxes,so I don't know how to use those images correctly.If I should label those images zero,and train those images as well as other images,or throw those images directly?</p>",
      "rawMarkdown": "I train an object detector based on [efficientdet](https://www.kaggle.com/shonenkov/timm-efficientdet-pytorch),but some images don't have bounding boxes,so I don't know how to use those images correctly.If I should label those images zero,and train those images as well as other images,or throw those images directly?",
      "votes": null
    },
    {
      "id": "1329604",
      "postDate": "05/31/2021 08:37:05",
      "content": "<p>Hello, I have the same question. Did you receive any answer to it? Please do let me know if you did. Anyways, I am <em>NOT</em> sure but what I am going to do is replace the NaN values with xmin = 0, ymin = 0, width = 1, and height = 1 since we have to predict none 1 0 0 1 1 if nothing is predicted. Again, I am not sure, I HAVE THE SAME QUESTION and I am a beginner. Cheers.</p>",
      "rawMarkdown": "Hello, I have the same question. Did you receive any answer to it? Please do let me know if you did. Anyways, I am *NOT* sure but what I am going to do is replace the NaN values with xmin = 0, ymin = 0, width = 1, and height = 1 since we have to predict none 1 0 0 1 1 if nothing is predicted. Again, I am not sure, I HAVE THE SAME QUESTION and I am a beginner. Cheers.",
      "votes": null
    },
    {
      "id": "1329791",
      "postDate": "05/31/2021 11:22:54",
      "content": "<p>Hi,thanks for your reply.I'm a beginner too.I will try using all images and only use images with bounding boxes after I debug my code.Actually I think these two ways will have similar results.You can try them on your self,or read object detector code directly,check how to compute loss when input images without bounding boxes.Good Luck!</p>",
      "rawMarkdown": "Hi,thanks for your reply.I'm a beginner too.I will try using all images and only use images with bounding boxes after I debug my code.Actually I think these two ways will have similar results.You can try them on your self,or read object detector code directly,check how to compute loss when input images without bounding boxes.Good Luck!",
      "votes": null
    },
    {
      "id": "1330629",
      "postDate": "06/01/2021 02:43:20",
      "content": "<p>Thanks for the idea hujianxin :D</p>",
      "rawMarkdown": "Thanks for the idea hujianxin :D",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1329604,
      "author_name": "eddwait",
      "author_url": "",
      "post_date": "05/31/2021 08:37:05",
      "content": "<p>Hello, I have the same question. Did you receive any answer to it? Please do let me know if you did. Anyways, I am <em>NOT</em> sure but what I am going to do is replace the NaN values with xmin = 0, ymin = 0, width = 1, and height = 1 since we have to predict none 1 0 0 1 1 if nothing is predicted. Again, I am not sure, I HAVE THE SAME QUESTION and I am a beginner. Cheers.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1329791,
          "author_name": "jianxinhu",
          "author_url": "",
          "post_date": "05/31/2021 11:22:54",
          "content": "<p>Hi,thanks for your reply.I'm a beginner too.I will try using all images and only use images with bounding boxes after I debug my code.Actually I think these two ways will have similar results.You can try them on your self,or read object detector code directly,check how to compute loss when input images without bounding boxes.Good Luck!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1330629,
          "author_name": "eddwait",
          "author_url": "",
          "post_date": "06/01/2021 02:43:20",
          "content": "<p>Thanks for the idea hujianxin :D</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1328485": "I train an object detector based on [efficientdet](https://www.kaggle.com/shonenkov/timm-efficientdet-pytorch),but some images don't have bounding boxes,so I don't know how to use those images correctly.If I should label those images zero,and train those images as well as other images,or throw those images directly?",
    "1329604": "Hello, I have the same question. Did you receive any answer to it? Please do let me know if you did. Anyways, I am *NOT* sure but what I am going to do is replace the NaN values with xmin = 0, ymin = 0, width = 1, and height = 1 since we have to predict none 1 0 0 1 1 if nothing is predicted. Again, I am not sure, I HAVE THE SAME QUESTION and I am a beginner. Cheers.",
    "1329791": "Hi,thanks for your reply.I'm a beginner too.I will try using all images and only use images with bounding boxes after I debug my code.Actually I think these two ways will have similar results.You can try them on your self,or read object detector code directly,check how to compute loss when input images without bounding boxes.Good Luck!",
    "1330629": "Thanks for the idea hujianxin :D"
  },
  "source": "meta"
}