{
  "id": 120381,
  "title": "Algorithm Selection in 6D Pose Estimation",
  "url": "/competitions/pku-autonomous-driving/discussion/120381",
  "author_name": "",
  "post_date": "2019-12-05T16:16:23.696145200Z",
  "votes": 11,
  "comment_count": 5,
  "views": 0,
  "content": "<p>This is summary of <a href=\"https://arxiv.org/abs/1905.06658v1\">Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning</a></p>\n\n<p>We can divided <strong>6D-Pose-Estimation Algorithm</strong> into 4 types.</p>\n\n<h2>1　Correspondence-based method ( high accuracy )</h2>\n\n<ul>\n<li>Project 3D model to RGB template image ( need 3D models )\n<ul><li>SIFT, SURF, ORB</li></ul></li>\n<li>3D to 3D ( need 3D point cloud, not fit this competition )\n<ul><li>Super 4PCS, SpinImages, FPFH, SHOT, ICP</li></ul></li>\n</ul>\n\n<h2>2　Template-based method ( limited accuracy )</h2>\n\n<ul>\n<li>lINEMOD</li>\n</ul>\n\n<h2>3　Voting-based method ( high accuracy but slow )</h2>\n\n<ul>\n<li>PPF, Object Coordinate, DenseFusion, PVNet</li>\n</ul>\n\n<h2>4　Regression-based method ( most public notebook use, need more datas )</h2>\n\n<ul>\n<li>Regress 6-D pose directly\n<ul><li>PoseCNN, Deep6DPose</li></ul></li>\n<li>Predict the projection position of key 3D points on 2D image, and then use PNP method to restore the pose\n<ul><li>BB8, SSD-6D, YOLO-6D</li></ul></li>\n</ul>\n\n<p>Not all the methods are suitable for this competition, the datas we own are <strong>RGB image</strong> and <strong>Car Models</strong>,  so google it to get more information.</p>\n\n<h3>If you think it's useful, please give me an upvote, thanks.</h3>\n\n<p>These are my other topics.\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120015#latest-686515\">Algorithm Selection for Beginner!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076\">The algorithm that Competition Organizer - baidu chooses!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785\">Understand Camera Intrinsic Parameters and X, Y, Z</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443\">How the Competition Organizer get train.csv !</a></p>\n\n<p>My notebook:\n<a href=\"https://www.kaggle.com/diegojohnson/centernet-objects-as-points\">Centernet - Objects as Points</a></p>",
  "messages": [
    {
      "id": "688477",
      "postDate": "12/05/2019 16:16:23",
      "content": "<p>This is summary of <a href=\"https://arxiv.org/abs/1905.06658v1\">Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning</a></p>\n\n<p>We can divided <strong>6D-Pose-Estimation Algorithm</strong> into 4 types.</p>\n\n<h2>1　Correspondence-based method ( high accuracy )</h2>\n\n<ul>\n<li>Project 3D model to RGB template image ( need 3D models )\n<ul><li>SIFT, SURF, ORB</li></ul></li>\n<li>3D to 3D ( need 3D point cloud, not fit this competition )\n<ul><li>Super 4PCS, SpinImages, FPFH, SHOT, ICP</li></ul></li>\n</ul>\n\n<h2>2　Template-based method ( limited accuracy )</h2>\n\n<ul>\n<li>lINEMOD</li>\n</ul>\n\n<h2>3　Voting-based method ( high accuracy but slow )</h2>\n\n<ul>\n<li>PPF, Object Coordinate, DenseFusion, PVNet</li>\n</ul>\n\n<h2>4　Regression-based method ( most public notebook use, need more datas )</h2>\n\n<ul>\n<li>Regress 6-D pose directly\n<ul><li>PoseCNN, Deep6DPose</li></ul></li>\n<li>Predict the projection position of key 3D points on 2D image, and then use PNP method to restore the pose\n<ul><li>BB8, SSD-6D, YOLO-6D</li></ul></li>\n</ul>\n\n<p>Not all the methods are suitable for this competition, the datas we own are <strong>RGB image</strong> and <strong>Car Models</strong>,  so google it to get more information.</p>\n\n<h3>If you think it's useful, please give me an upvote, thanks.</h3>\n\n<p>These are my other topics.\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120015#latest-686515\">Algorithm Selection for Beginner!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076\">The algorithm that Competition Organizer - baidu chooses!</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785\">Understand Camera Intrinsic Parameters and X, Y, Z</a>\n<a href=\"https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443\">How the Competition Organizer get train.csv !</a></p>\n\n<p>My notebook:\n<a href=\"https://www.kaggle.com/diegojohnson/centernet-objects-as-points\">Centernet - Objects as Points</a></p>",
      "rawMarkdown": "This is summary of [Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning](https://arxiv.org/abs/1905.06658v1)\n\nWe can divided **6D-Pose-Estimation Algorithm** into 4 types.\n\n## 1　Correspondence-based method ( high accuracy )\n- Project 3D model to RGB template image ( need 3D models )\n *  SIFT, SURF, ORB\n- 3D to 3D ( need 3D point cloud, not fit this competition )\n *  Super 4PCS, SpinImages, FPFH, SHOT, ICP\n\n## 2　Template-based method ( limited accuracy )\n * lINEMOD\n\n## 3　Voting-based method ( high accuracy but slow )\n- PPF, Object Coordinate, DenseFusion, PVNet\n\n\n## 4　Regression-based method ( most public notebook use, need more datas )\n- Regress 6-D pose directly\n * PoseCNN, Deep6DPose\n- Predict the projection position of key 3D points on 2D image, and then use PNP method to restore the pose\n * BB8, SSD-6D, YOLO-6D\n\nNot all the methods are suitable for this competition, the datas we own are **RGB image** and **Car Models**,  so google it to get more information.\n\n### If you think it's useful, please give me an upvote, thanks.\n\nThese are my other topics.\n[Algorithm Selection for Beginner!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120015#latest-686515)\n[The algorithm that Competition Organizer - baidu chooses!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076)\n[Understand Camera Intrinsic Parameters and X, Y, Z](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785)\n[How the Competition Organizer get train.csv !](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443)\n\nMy notebook:\n[Centernet - Objects as Points](https://www.kaggle.com/diegojohnson/centernet-objects-as-points)",
      "votes": null
    },
    {
      "id": "688866",
      "postDate": "12/06/2019 07:24:55",
      "content": "<p>I haven't see anybody in public notebook using car-models in their algorithm. How to use it, anybody can give some suggestions ?</p>",
      "rawMarkdown": "I haven't see anybody in public notebook using car-models in their algorithm. How to use it, anybody can give some suggestions ?",
      "votes": null
    },
    {
      "id": "692158",
      "postDate": "12/11/2019 01:14:30",
      "content": "<p>I have the same question no one using 3d car models. I think car models can b used for getting yaw, pitch, and roll.</p>",
      "rawMarkdown": "I have the same question no one using 3d car models. I think car models can b used for getting yaw, pitch, and roll.",
      "votes": null
    },
    {
      "id": "692171",
      "postDate": "12/11/2019 01:36:32",
      "content": "<p>You are right, but no one share their method.\nI think the best way is fine-tuning rotations with 3d models after regression.</p>",
      "rawMarkdown": "You are right, but no one share their method.\nI think the best way is fine-tuning rotations with 3d models after regression.",
      "votes": null
    },
    {
      "id": "692176",
      "postDate": "12/11/2019 01:41:06",
      "content": "<p>I think first we find the car id and location and then use the car model by putting it on its location. It is just a theory. 😄 </p>",
      "rawMarkdown": "I think first we find the car id and location and then use the car model by putting it on its location. It is just a theory. 😄",
      "votes": null
    },
    {
      "id": "704816",
      "postDate": "12/28/2019 03:19:52",
      "content": "<p>I'm wonder why we choose the algorithms that the Kitti-Competition choose? \nJust like MonoDis, Roi-10D...</p>",
      "rawMarkdown": "I'm wonder why we choose the algorithms that the Kitti-Competition choose? \nJust like MonoDis, Roi-10D...",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 688866,
      "author_name": "diegojohnson",
      "author_url": "",
      "post_date": "12/06/2019 07:24:55",
      "content": "<p>I haven't see anybody in public notebook using car-models in their algorithm. How to use it, anybody can give some suggestions ?</p>",
      "votes": null,
      "replies": [
        {
          "id": 692158,
          "author_name": "akmalmasud96",
          "author_url": "",
          "post_date": "12/11/2019 01:14:30",
          "content": "<p>I have the same question no one using 3d car models. I think car models can b used for getting yaw, pitch, and roll.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 692171,
          "author_name": "diegojohnson",
          "author_url": "",
          "post_date": "12/11/2019 01:36:32",
          "content": "<p>You are right, but no one share their method.\nI think the best way is fine-tuning rotations with 3d models after regression.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 692176,
          "author_name": "akmalmasud96",
          "author_url": "",
          "post_date": "12/11/2019 01:41:06",
          "content": "<p>I think first we find the car id and location and then use the car model by putting it on its location. It is just a theory. 😄 </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 704816,
      "author_name": "johnzdh",
      "author_url": "",
      "post_date": "12/28/2019 03:19:52",
      "content": "<p>I'm wonder why we choose the algorithms that the Kitti-Competition choose? \nJust like MonoDis, Roi-10D...</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "688477": "This is summary of [Vision-based Robotic Grasping from Object Localization, Pose Estimation, Grasp Detection to Motion Planning](https://arxiv.org/abs/1905.06658v1)\n\nWe can divided **6D-Pose-Estimation Algorithm** into 4 types.\n\n## 1　Correspondence-based method ( high accuracy )\n- Project 3D model to RGB template image ( need 3D models )\n *  SIFT, SURF, ORB\n- 3D to 3D ( need 3D point cloud, not fit this competition )\n *  Super 4PCS, SpinImages, FPFH, SHOT, ICP\n\n## 2　Template-based method ( limited accuracy )\n * lINEMOD\n\n## 3　Voting-based method ( high accuracy but slow )\n- PPF, Object Coordinate, DenseFusion, PVNet\n\n\n## 4　Regression-based method ( most public notebook use, need more datas )\n- Regress 6-D pose directly\n * PoseCNN, Deep6DPose\n- Predict the projection position of key 3D points on 2D image, and then use PNP method to restore the pose\n * BB8, SSD-6D, YOLO-6D\n\nNot all the methods are suitable for this competition, the datas we own are **RGB image** and **Car Models**,  so google it to get more information.\n\n### If you think it's useful, please give me an upvote, thanks.\n\nThese are my other topics.\n[Algorithm Selection for Beginner!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120015#latest-686515)\n[The algorithm that Competition Organizer - baidu chooses!](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120076)\n[Understand Camera Intrinsic Parameters and X, Y, Z](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120083#latest-686785)\n[How the Competition Organizer get train.csv !](https://www.kaggle.com/c/pku-autonomous-driving/discussion/120443)\n\nMy notebook:\n[Centernet - Objects as Points](https://www.kaggle.com/diegojohnson/centernet-objects-as-points)",
    "688866": "I haven't see anybody in public notebook using car-models in their algorithm. How to use it, anybody can give some suggestions ?",
    "692158": "I have the same question no one using 3d car models. I think car models can b used for getting yaw, pitch, and roll.",
    "692171": "You are right, but no one share their method.\nI think the best way is fine-tuning rotations with 3d models after regression.",
    "692176": "I think first we find the car id and location and then use the car model by putting it on its location. It is just a theory. 😄",
    "704816": "I'm wonder why we choose the algorithms that the Kitti-Competition choose? \nJust like MonoDis, Roi-10D..."
  },
  "source": "meta"
}