{
  "id": 294794,
  "title": "Using previous frame's annotations (Kalman filter)?",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/294794",
  "author_name": "",
  "post_date": "2021-12-12T20:31:15.377060100Z",
  "votes": 7,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Is anyone working on updating the annotation of frame <code>i</code> by incorporating annotations from frame <code>i-n</code>? But of course the location (and size) of the COTS moves as the camera moves forward and over the reef, so creating \"tubelets\" for the annotations is not trivial. But two thoughts would be:</p>\n<ul>\n<li>a <a href=\"https://www.bzarg.com/p/how-a-kalman-filter-works-in-pictures/\" target=\"_blank\">Kalman filter</a>; or</li>\n<li>a sequence prediction model (LSTM or XGBoost or somesuch)</li>\n</ul>\n<p>I'm thinking that this would be separate from the main object detection model, so that it could potentially give a boost to any single-frame detection model. </p>",
  "messages": [
    {
      "id": "1615877",
      "postDate": "12/12/2021 20:31:15",
      "content": "<p>Is anyone working on updating the annotation of frame <code>i</code> by incorporating annotations from frame <code>i-n</code>? But of course the location (and size) of the COTS moves as the camera moves forward and over the reef, so creating \"tubelets\" for the annotations is not trivial. But two thoughts would be:</p>\n<ul>\n<li>a <a href=\"https://www.bzarg.com/p/how-a-kalman-filter-works-in-pictures/\" target=\"_blank\">Kalman filter</a>; or</li>\n<li>a sequence prediction model (LSTM or XGBoost or somesuch)</li>\n</ul>\n<p>I'm thinking that this would be separate from the main object detection model, so that it could potentially give a boost to any single-frame detection model. </p>",
      "rawMarkdown": "Is anyone working on updating the annotation of frame `i` by incorporating annotations from frame `i-n`? But of course the location (and size) of the COTS moves as the camera moves forward and over the reef, so creating \"tubelets\" for the annotations is not trivial. But two thoughts would be:\n\n- a [Kalman filter](https://www.bzarg.com/p/how-a-kalman-filter-works-in-pictures/); or\n- a sequence prediction model (LSTM or XGBoost or somesuch)\n\nI'm thinking that this would be separate from the main object detection model, so that it could potentially give a boost to any single-frame detection model.",
      "votes": null
    },
    {
      "id": "1617849",
      "postDate": "12/14/2021 12:32:40",
      "content": "<p>for tracking the bboxes we can use DeepSORT, which was greatly used in the previous NFL competition for tracking the helmets in the game.<br>\n<a href=\"https://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort\" target=\"_blank\">https://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort</a></p>",
      "rawMarkdown": "for tracking the bboxes we can use DeepSORT, which was greatly used in the previous NFL competition for tracking the helmets in the game.\nhttps://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1617849,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "12/14/2021 12:32:40",
      "content": "<p>for tracking the bboxes we can use DeepSORT, which was greatly used in the previous NFL competition for tracking the helmets in the game.<br>\n<a href=\"https://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort\" target=\"_blank\">https://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1615877": "Is anyone working on updating the annotation of frame `i` by incorporating annotations from frame `i-n`? But of course the location (and size) of the COTS moves as the camera moves forward and over the reef, so creating \"tubelets\" for the annotations is not trivial. But two thoughts would be:\n\n- a [Kalman filter](https://www.bzarg.com/p/how-a-kalman-filter-works-in-pictures/); or\n- a sequence prediction model (LSTM or XGBoost or somesuch)\n\nI'm thinking that this would be separate from the main object detection model, so that it could potentially give a boost to any single-frame detection model.",
    "1617849": "for tracking the bboxes we can use DeepSORT, which was greatly used in the previous NFL competition for tracking the helmets in the game.\nhttps://www.kaggle.com/robikscube/helper-code-helmet-mapping-deepsort"
  },
  "source": "meta"
}