{
  "id": 186495,
  "title": "Agent Positions from Zarr File",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/186495",
  "author_name": "",
  "post_date": "2020-09-24T17:15:11.866285900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I want to experiment with prediction using agent positions and want more flexibility (and less processing time) than I can get straight through the Ego/AgentDataset.</p>\n<p>I have no problem grabbing this data by picking a scene, grabbing the frames from the scene, and then grabbing the track_ids that appear in those frames.   That should allow me to monitor a specific agent as it moves through the frames.  </p>\n<p>However, when I try to bounce what I am getting using this technique against what I get using the built in Ego/AgentDataset it doesn't seem like I see the same data (for example, track_ids tend to  appear in more frames using my method than using the AgentDataset method, the same set of track_ids don't seem to appear in the same scenes, etc).  </p>\n<p>Is there any underlying issue with my approach or I am comparing apples to oranges trying to correlate the two methods? </p>",
  "messages": [
    {
      "id": "1025615",
      "postDate": "09/24/2020 17:15:11",
      "content": "<p>I want to experiment with prediction using agent positions and want more flexibility (and less processing time) than I can get straight through the Ego/AgentDataset.</p>\n<p>I have no problem grabbing this data by picking a scene, grabbing the frames from the scene, and then grabbing the track_ids that appear in those frames.   That should allow me to monitor a specific agent as it moves through the frames.  </p>\n<p>However, when I try to bounce what I am getting using this technique against what I get using the built in Ego/AgentDataset it doesn't seem like I see the same data (for example, track_ids tend to  appear in more frames using my method than using the AgentDataset method, the same set of track_ids don't seem to appear in the same scenes, etc).  </p>\n<p>Is there any underlying issue with my approach or I am comparing apples to oranges trying to correlate the two methods? </p>",
      "rawMarkdown": "I want to experiment with prediction using agent positions and want more flexibility (and less processing time) than I can get straight through the Ego/AgentDataset.\n\nI have no problem grabbing this data by picking a scene, grabbing the frames from the scene, and then grabbing the track_ids that appear in those frames.   That should allow me to monitor a specific agent as it moves through the frames.  \n\nHowever, when I try to bounce what I am getting using this technique against what I get using the built in Ego/AgentDataset it doesn't seem like I see the same data (for example, track_ids tend to  appear in more frames using my method than using the AgentDataset method, the same set of track_ids don't seem to appear in the same scenes, etc).  \n\nIs there any underlying issue with my approach or I am comparing apples to oranges trying to correlate the two methods?",
      "votes": null
    },
    {
      "id": "1025790",
      "postDate": "09/24/2020 19:07:30",
      "content": "<blockquote>\n  <p>for example, track_ids tend to appear in more frames using my method than using the AgentDataset method  </p>\n</blockquote>\n<p>This could be due to the fact that the AgentDataset will only have agents with at least 10 frames of history (try setting <code>min_frame_history = 0</code>), so it will only be included in the AgentDataset once it ha been visible for at least 10 frames</p>",
      "rawMarkdown": "> for example, track_ids tend to appear in more frames using my method than using the AgentDataset method  \n\nThis could be due to the fact that the AgentDataset will only have agents with at least 10 frames of history (try setting `min_frame_history = 0`), so it will only be included in the AgentDataset once it ha been visible for at least 10 frames",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1025790,
      "author_name": "fnands",
      "author_url": "",
      "post_date": "09/24/2020 19:07:30",
      "content": "<blockquote>\n  <p>for example, track_ids tend to appear in more frames using my method than using the AgentDataset method  </p>\n</blockquote>\n<p>This could be due to the fact that the AgentDataset will only have agents with at least 10 frames of history (try setting <code>min_frame_history = 0</code>), so it will only be included in the AgentDataset once it ha been visible for at least 10 frames</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1025615": "I want to experiment with prediction using agent positions and want more flexibility (and less processing time) than I can get straight through the Ego/AgentDataset.\n\nI have no problem grabbing this data by picking a scene, grabbing the frames from the scene, and then grabbing the track_ids that appear in those frames.   That should allow me to monitor a specific agent as it moves through the frames.  \n\nHowever, when I try to bounce what I am getting using this technique against what I get using the built in Ego/AgentDataset it doesn't seem like I see the same data (for example, track_ids tend to  appear in more frames using my method than using the AgentDataset method, the same set of track_ids don't seem to appear in the same scenes, etc).  \n\nIs there any underlying issue with my approach or I am comparing apples to oranges trying to correlate the two methods?",
    "1025790": "> for example, track_ids tend to appear in more frames using my method than using the AgentDataset method  \n\nThis could be due to the fact that the AgentDataset will only have agents with at least 10 frames of history (try setting `min_frame_history = 0`), so it will only be included in the AgentDataset once it ha been visible for at least 10 frames"
  },
  "source": "meta"
}