{
  "id": 115868,
  "title": "Lyft to KITTI convertor: what happens to non-front cameras?",
  "url": "/competitions/3d-object-detection-for-autonomous-vehicles/discussion/115868",
  "author_name": "Artyom Palvelev",
  "post_date": "2019-11-05T18:45:09.332000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I'm using F-ConvNet and a KITTI converter from Lyft SDK. I've got relatively decent predictions for the <code>CAM_FRONT</code>, but my model's predictions are wrong for all other cameras. It returns something remotely similar for <code>CAM_LEFT/RIGHT</code> and basically garbage for other cameras (cars are deep under the ground).</p>\n\n<p>Am I right when I'm thinking that KITTI converter returns both point cloud and boxes in Lidar coordinate system? Do I need any transformations for non-default cameras?</p>\n\n<p>I understand that F-ConvNet centers the point cloud against frustum's axis, so it should work. It looks like a coordinate space problem to me.</p>\n\n<p><code>CAM_FRONT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F2f55760c8e39ea8ad3ff9e1bd02886c6%2Fcam_front.png?generation=1572979283889645&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_FRONT_RIGHT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F07deffd16a14b3128b1d8b8e5888a52c%2Fcam_right.png?generation=1572979319155709&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_BACK_LEFT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F6098661dced10e76c09d29308b5040e9%2Fcam_back_left.png?generation=1572979374577461&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_BACK</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2Fbae747bf0b1366d7f22c3719e826b0cd%2Fcam_back.png?generation=1572979353765494&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": 666121,
      "postDate": "2019-11-05T18:45:09.333Z",
      "content": "<p>I'm using F-ConvNet and a KITTI converter from Lyft SDK. I've got relatively decent predictions for the <code>CAM_FRONT</code>, but my model's predictions are wrong for all other cameras. It returns something remotely similar for <code>CAM_LEFT/RIGHT</code> and basically garbage for other cameras (cars are deep under the ground).</p>\n\n<p>Am I right when I'm thinking that KITTI converter returns both point cloud and boxes in Lidar coordinate system? Do I need any transformations for non-default cameras?</p>\n\n<p>I understand that F-ConvNet centers the point cloud against frustum's axis, so it should work. It looks like a coordinate space problem to me.</p>\n\n<p><code>CAM_FRONT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F2f55760c8e39ea8ad3ff9e1bd02886c6%2Fcam_front.png?generation=1572979283889645&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_FRONT_RIGHT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F07deffd16a14b3128b1d8b8e5888a52c%2Fcam_right.png?generation=1572979319155709&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_BACK_LEFT</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F6098661dced10e76c09d29308b5040e9%2Fcam_back_left.png?generation=1572979374577461&amp;alt=media\" alt=\"\"></p>\n\n<p><code>CAM_BACK</code>:\n<img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2Fbae747bf0b1366d7f22c3719e826b0cd%2Fcam_back.png?generation=1572979353765494&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "I'm using F-ConvNet and a KITTI converter from Lyft SDK. I've got relatively decent predictions for the `CAM_FRONT`, but my model's predictions are wrong for all other cameras. It returns something remotely similar for `CAM_LEFT/RIGHT` and basically garbage for other cameras (cars are deep under the ground).\n\nAm I right when I'm thinking that KITTI converter returns both point cloud and boxes in Lidar coordinate system? Do I need any transformations for non-default cameras?\n\nI understand that F-ConvNet centers the point cloud against frustum's axis, so it should work. It looks like a coordinate space problem to me.\n\n`CAM_FRONT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F2f55760c8e39ea8ad3ff9e1bd02886c6%2Fcam_front.png?generation=1572979283889645&amp;alt=media)\n\n`CAM_FRONT_RIGHT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F07deffd16a14b3128b1d8b8e5888a52c%2Fcam_right.png?generation=1572979319155709&amp;alt=media)\n\n`CAM_BACK_LEFT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F6098661dced10e76c09d29308b5040e9%2Fcam_back_left.png?generation=1572979374577461&amp;alt=media)\n\n\n`CAM_BACK`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2Fbae747bf0b1366d7f22c3719e826b0cd%2Fcam_back.png?generation=1572979353765494&amp;alt=media)\n",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "666121": "I'm using F-ConvNet and a KITTI converter from Lyft SDK. I've got relatively decent predictions for the `CAM_FRONT`, but my model's predictions are wrong for all other cameras. It returns something remotely similar for `CAM_LEFT/RIGHT` and basically garbage for other cameras (cars are deep under the ground).\n\nAm I right when I'm thinking that KITTI converter returns both point cloud and boxes in Lidar coordinate system? Do I need any transformations for non-default cameras?\n\nI understand that F-ConvNet centers the point cloud against frustum's axis, so it should work. It looks like a coordinate space problem to me.\n\n`CAM_FRONT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F2f55760c8e39ea8ad3ff9e1bd02886c6%2Fcam_front.png?generation=1572979283889645&amp;alt=media)\n\n`CAM_FRONT_RIGHT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F07deffd16a14b3128b1d8b8e5888a52c%2Fcam_right.png?generation=1572979319155709&amp;alt=media)\n\n`CAM_BACK_LEFT`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2F6098661dced10e76c09d29308b5040e9%2Fcam_back_left.png?generation=1572979374577461&amp;alt=media)\n\n\n`CAM_BACK`:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1859557%2Fbae747bf0b1366d7f22c3719e826b0cd%2Fcam_back.png?generation=1572979353765494&amp;alt=media)\n"
  }
}