{
  "id": 115479,
  "title": "Ideas of using 3D mode",
  "url": "/competitions/pku-autonomous-driving/discussion/115479",
  "author_name": "",
  "post_date": "2019-11-03T03:29:25.217254200Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>There are some Notebooks show us how to visualize the 3D model, but so far I haven't seen people share how to use 3D model to get better prediction.</p>\n\n<p>Here is what I think 3D model may help :</p>\n\n<ol>\n<li>Using any kind of model to predict the 7 value in the image (typically using Centernet, cause we don't have length and height for car in the image)</li>\n<li>To use 3D model, we can use these 7 value as input, and add car's height, length,depth calculated by 3D model, plus semantic from Centernet to further regress(refine) the prediction.</li>\n</ol>\n\n<p>Anybody have idea about how to use 3D model?</p>",
  "messages": [
    {
      "id": "664017",
      "postDate": "11/03/2019 03:29:25",
      "content": "<p>There are some Notebooks show us how to visualize the 3D model, but so far I haven't seen people share how to use 3D model to get better prediction.</p>\n\n<p>Here is what I think 3D model may help :</p>\n\n<ol>\n<li>Using any kind of model to predict the 7 value in the image (typically using Centernet, cause we don't have length and height for car in the image)</li>\n<li>To use 3D model, we can use these 7 value as input, and add car's height, length,depth calculated by 3D model, plus semantic from Centernet to further regress(refine) the prediction.</li>\n</ol>\n\n<p>Anybody have idea about how to use 3D model?</p>",
      "rawMarkdown": "There are some Notebooks show us how to visualize the 3D model, but so far I haven't seen people share how to use 3D model to get better prediction.\n\nHere is what I think 3D model may help :\n\n1. Using any kind of model to predict the 7 value in the image (typically using Centernet, cause we don't have length and height for car in the image)\n2. To use 3D model, we can use these 7 value as input, and add car's height, length,depth calculated by 3D model, plus semantic from Centernet to further regress(refine) the prediction.\n\nAnybody have idea about how to use 3D model?",
      "votes": null
    },
    {
      "id": "670524",
      "postDate": "11/11/2019 14:48:28",
      "content": "<p>I haven't looked at the data in detail yet, but the first thing that comes to mind is that 3D models can be used for 3D augmentation and generation of synthetic data, where you can debug scripts or pre-train some parts of the model.</p>",
      "rawMarkdown": "I haven't looked at the data in detail yet, but the first thing that comes to mind is that 3D models can be used for 3D augmentation and generation of synthetic data, where you can debug scripts or pre-train some parts of the model.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 670524,
      "author_name": "goodok",
      "author_url": "",
      "post_date": "11/11/2019 14:48:28",
      "content": "<p>I haven't looked at the data in detail yet, but the first thing that comes to mind is that 3D models can be used for 3D augmentation and generation of synthetic data, where you can debug scripts or pre-train some parts of the model.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "664017": "There are some Notebooks show us how to visualize the 3D model, but so far I haven't seen people share how to use 3D model to get better prediction.\n\nHere is what I think 3D model may help :\n\n1. Using any kind of model to predict the 7 value in the image (typically using Centernet, cause we don't have length and height for car in the image)\n2. To use 3D model, we can use these 7 value as input, and add car's height, length,depth calculated by 3D model, plus semantic from Centernet to further regress(refine) the prediction.\n\nAnybody have idea about how to use 3D model?",
    "670524": "I haven't looked at the data in detail yet, but the first thing that comes to mind is that 3D models can be used for 3D augmentation and generation of synthetic data, where you can debug scripts or pre-train some parts of the model."
  },
  "source": "meta"
}