{
  "id": 302779,
  "title": "Leveraging temporal information",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/302779",
  "author_name": "",
  "post_date": "2022-01-24T09:05:27.727968600Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>Hello everyone, <br>\nIt is my first object detection competition and I've been working on this problem mainly to learn about detection problematics (BBox , metrics etc…).<br>\nThis competition is not object detection in images only but images coming from the same videos though. Most notebooks I've made or seen here so far are approaching the problem by applying classical transformation and pipelines to the images - do you think it would be beneficial to approach this problem as a time series problem and leveraging the fact that images come in a specific order due to being from the same videos ?</p>\n<p>Thanks, and good luck to all in this competition</p>",
  "messages": [
    {
      "id": "1662416",
      "postDate": "01/24/2022 09:05:27",
      "content": "<p>Hello everyone, <br>\nIt is my first object detection competition and I've been working on this problem mainly to learn about detection problematics (BBox , metrics etc…).<br>\nThis competition is not object detection in images only but images coming from the same videos though. Most notebooks I've made or seen here so far are approaching the problem by applying classical transformation and pipelines to the images - do you think it would be beneficial to approach this problem as a time series problem and leveraging the fact that images come in a specific order due to being from the same videos ?</p>\n<p>Thanks, and good luck to all in this competition</p>",
      "rawMarkdown": "Hello everyone, \nIt is my first object detection competition and I've been working on this problem mainly to learn about detection problematics (BBox , metrics etc...).\nThis competition is not object detection in images only but images coming from the same videos though. Most notebooks I've made or seen here so far are approaching the problem by applying classical transformation and pipelines to the images - do you think it would be beneficial to approach this problem as a time series problem and leveraging the fact that images come in a specific order due to being from the same videos ?\n\nThanks, and good luck to all in this competition",
      "votes": null
    },
    {
      "id": "1662497",
      "postDate": "01/24/2022 10:12:53",
      "content": "<p>I think both approaches are important.</p>\n<p>On one end you want to have a robust model that can identify the possible starfish in a given image, but you also want to take into account the temporal component, that a starfish present on a frame x will likely be present in frame x+1.</p>",
      "rawMarkdown": "I think both approaches are important.\n\nOn one end you want to have a robust model that can identify the possible starfish in a given image, but you also want to take into account the temporal component, that a starfish present on a frame x will likely be present in frame x+1.",
      "votes": null
    },
    {
      "id": "1662678",
      "postDate": "01/24/2022 13:12:46",
      "content": "<p>I agree with you, but I'm not able to find a way to use temp info, I mean, do you know some trick that we can use to get it without need to use a lot more resources? I was thinking to try a RNN using a output from a middle layer of my current model, but I'm failing to find how to do it…</p>",
      "rawMarkdown": "I agree with you, but I'm not able to find a way to use temp info, I mean, do you know some trick that we can use to get it without need to use a lot more resources? I was thinking to try a RNN using a output from a middle layer of my current model, but I'm failing to find how to do it...",
      "votes": null
    },
    {
      "id": "1662889",
      "postDate": "01/24/2022 16:06:17",
      "content": "<p>Thanks for your answer, just like HABP I would like to know more - do you have any ressources pointing to the use of a temporal component when doing object detection or similar tasks ?</p>",
      "rawMarkdown": "Thanks for your answer, just like HABP I would like to know more - do you have any ressources pointing to the use of a temporal component when doing object detection or similar tasks ?",
      "votes": null
    },
    {
      "id": "1662919",
      "postDate": "01/24/2022 16:41:45",
      "content": "<p>That is a good question I haven't found a good answer for yet (I am quite new to Computer Vision). </p>\n<p>If you want something cheap and with, you can try inference with tracking. It might help a bit, but it is not learning the temporarity directly.</p>\n<p>Regarding specific models, there are some really interesting papers in the domain of Video Object Detection that might be worth looking into. I am trying to implement a couple ideas from there.</p>",
      "rawMarkdown": "That is a good question I haven't found a good answer for yet (I am quite new to Computer Vision). \n\nIf you want something cheap and with, you can try inference with tracking. It might help a bit, but it is not learning the temporarity directly.\n\nRegarding specific models, there are some really interesting papers in the domain of Video Object Detection that might be worth looking into. I am trying to implement a couple ideas from there.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1662497,
      "author_name": "dauriel",
      "author_url": "",
      "post_date": "01/24/2022 10:12:53",
      "content": "<p>I think both approaches are important.</p>\n<p>On one end you want to have a robust model that can identify the possible starfish in a given image, but you also want to take into account the temporal component, that a starfish present on a frame x will likely be present in frame x+1.</p>",
      "votes": null,
      "replies": [
        {
          "id": 1662678,
          "author_name": "henriqueabpassos",
          "author_url": "",
          "post_date": "01/24/2022 13:12:46",
          "content": "<p>I agree with you, but I'm not able to find a way to use temp info, I mean, do you know some trick that we can use to get it without need to use a lot more resources? I was thinking to try a RNN using a output from a middle layer of my current model, but I'm failing to find how to do it…</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1662889,
          "author_name": "djilax",
          "author_url": "",
          "post_date": "01/24/2022 16:06:17",
          "content": "<p>Thanks for your answer, just like HABP I would like to know more - do you have any ressources pointing to the use of a temporal component when doing object detection or similar tasks ?</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1662919,
      "author_name": "dauriel",
      "author_url": "",
      "post_date": "01/24/2022 16:41:45",
      "content": "<p>That is a good question I haven't found a good answer for yet (I am quite new to Computer Vision). </p>\n<p>If you want something cheap and with, you can try inference with tracking. It might help a bit, but it is not learning the temporarity directly.</p>\n<p>Regarding specific models, there are some really interesting papers in the domain of Video Object Detection that might be worth looking into. I am trying to implement a couple ideas from there.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1662416": "Hello everyone, \nIt is my first object detection competition and I've been working on this problem mainly to learn about detection problematics (BBox , metrics etc...).\nThis competition is not object detection in images only but images coming from the same videos though. Most notebooks I've made or seen here so far are approaching the problem by applying classical transformation and pipelines to the images - do you think it would be beneficial to approach this problem as a time series problem and leveraging the fact that images come in a specific order due to being from the same videos ?\n\nThanks, and good luck to all in this competition",
    "1662497": "I think both approaches are important.\n\nOn one end you want to have a robust model that can identify the possible starfish in a given image, but you also want to take into account the temporal component, that a starfish present on a frame x will likely be present in frame x+1.",
    "1662678": "I agree with you, but I'm not able to find a way to use temp info, I mean, do you know some trick that we can use to get it without need to use a lot more resources? I was thinking to try a RNN using a output from a middle layer of my current model, but I'm failing to find how to do it...",
    "1662889": "Thanks for your answer, just like HABP I would like to know more - do you have any ressources pointing to the use of a temporal component when doing object detection or similar tasks ?",
    "1662919": "That is a good question I haven't found a good answer for yet (I am quite new to Computer Vision). \n\nIf you want something cheap and with, you can try inference with tracking. It might help a bit, but it is not learning the temporarity directly.\n\nRegarding specific models, there are some really interesting papers in the domain of Video Object Detection that might be worth looking into. I am trying to implement a couple ideas from there."
  },
  "source": "meta"
}