{
  "id": 162642,
  "title": "Object Detection with Facebook's Detection Transformer",
  "url": "/competitions/global-wheat-detection/discussion/162642",
  "author_name": "",
  "post_date": "2020-06-29T15:51:32.762793800Z",
  "votes": 10,
  "comment_count": 7,
  "views": 0,
  "content": "<p>Ever since their inception transformers, the main research has been around trying to leverage them in different tasks apart from NLP. Facebook research launched a recent paper END TO END OBJECT DETECTION USING TRANSFORMERS :DETR. It uses transformers for Detection of objects using a new loss function. I have written a pipeline to fine-tune DETR on custom dataset using this competition to achieve state of the art results. </p>\n\n<p><a href=\"https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\">https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr</a></p>\n\n<p>DETR gives a fairly good result with training on a single epoch itself , being a transformer model , it does not require a lot of epoch to train, also with lr strategy. I am planning to update this kernel after complete training and also more kernels on this topic.</p>",
  "messages": [
    {
      "id": "906904",
      "postDate": "06/29/2020 15:51:32",
      "content": "<p>Ever since their inception transformers, the main research has been around trying to leverage them in different tasks apart from NLP. Facebook research launched a recent paper END TO END OBJECT DETECTION USING TRANSFORMERS :DETR. It uses transformers for Detection of objects using a new loss function. I have written a pipeline to fine-tune DETR on custom dataset using this competition to achieve state of the art results. </p>\n\n<p><a href=\"https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\">https://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr</a></p>\n\n<p>DETR gives a fairly good result with training on a single epoch itself , being a transformer model , it does not require a lot of epoch to train, also with lr strategy. I am planning to update this kernel after complete training and also more kernels on this topic.</p>",
      "rawMarkdown": "Ever since their inception transformers, the main research has been around trying to leverage them in different tasks apart from NLP. Facebook research launched a recent paper END TO END OBJECT DETECTION USING TRANSFORMERS :DETR. It uses transformers for Detection of objects using a new loss function. I have written a pipeline to fine-tune DETR on custom dataset using this competition to achieve state of the art results. \n\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\n\nDETR gives a fairly good result with training on a single epoch itself , being a transformer model , it does not require a lot of epoch to train, also with lr strategy. I am planning to update this kernel after complete training and also more kernels on this topic.",
      "votes": null
    },
    {
      "id": "906924",
      "postDate": "06/29/2020 16:03:33",
      "content": "<p>Great notebook! I would like to ask you one question, can you tell how good it performs as compared to YOLOv5?</p>",
      "rawMarkdown": "Great notebook! I would like to ask you one question, can you tell how good it performs as compared to YOLOv5?",
      "votes": null
    },
    {
      "id": "907040",
      "postDate": "06/29/2020 16:59:54",
      "content": "<p>Great work!</p>",
      "rawMarkdown": "Great work!",
      "votes": null
    },
    {
      "id": "907124",
      "postDate": "06/29/2020 17:37:27",
      "content": "<p>I am still in early stages and I have not fully started with fine tuning yet, I will start tomorrow after m5 , maybe then I will update results here</p>",
      "rawMarkdown": "I am still in early stages and I have not fully started with fine tuning yet, I will start tomorrow after m5 , maybe then I will update results here",
      "votes": null
    },
    {
      "id": "907152",
      "postDate": "06/29/2020 17:58:59",
      "content": "<p>Cool! Waiting to see the benchmarks :)</p>",
      "rawMarkdown": "Cool! Waiting to see the benchmarks :)",
      "votes": null
    },
    {
      "id": "907154",
      "postDate": "06/29/2020 18:02:04",
      "content": "<p>Thanks <a href=\"/anmspro\">@anmspro</a> </p>",
      "rawMarkdown": "Thanks @anmspro",
      "votes": null
    },
    {
      "id": "907204",
      "postDate": "06/29/2020 18:56:58",
      "content": "<p>DETR is very interesting , I heard it's very good for segmentation too.</p>",
      "rawMarkdown": "DETR is very interesting , I heard it's very good for segmentation too.",
      "votes": null
    },
    {
      "id": "963268",
      "postDate": "08/08/2020 20:14:47",
      "content": "<p>Great piece of work! I have tried to re implement the DeTr with some changement in the minimal code they have shared in the original paper. Unfortunatly my re implementation do not work well, I think I know where is located the issue since I am not using the NestedTensor (I think they do the absraction for no object labels )in my architecture. Do someone have tried to do the same experience?</p>",
      "rawMarkdown": "Great piece of work! I have tried to re implement the DeTr with some changement in the minimal code they have shared in the original paper. Unfortunatly my re implementation do not work well, I think I know where is located the issue since I am not using the NestedTensor (I think they do the absraction for no object labels )in my architecture. Do someone have tried to do the same experience?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 963268,
      "author_name": "rachidyz",
      "author_url": "",
      "post_date": "08/08/2020 20:14:47",
      "content": "<p>Great piece of work! I have tried to re implement the DeTr with some changement in the minimal code they have shared in the original paper. Unfortunatly my re implementation do not work well, I think I know where is located the issue since I am not using the NestedTensor (I think they do the absraction for no object labels )in my architecture. Do someone have tried to do the same experience?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 906924,
      "author_name": "aadhavvignesh",
      "author_url": "",
      "post_date": "06/29/2020 16:03:33",
      "content": "<p>Great notebook! I would like to ask you one question, can you tell how good it performs as compared to YOLOv5?</p>",
      "votes": null,
      "replies": [
        {
          "id": 907124,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "06/29/2020 17:37:27",
          "content": "<p>I am still in early stages and I have not fully started with fine tuning yet, I will start tomorrow after m5 , maybe then I will update results here</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 907152,
          "author_name": "aadhavvignesh",
          "author_url": "",
          "post_date": "06/29/2020 17:58:59",
          "content": "<p>Cool! Waiting to see the benchmarks :)</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 907040,
      "author_name": "anmspro",
      "author_url": "",
      "post_date": "06/29/2020 16:59:54",
      "content": "<p>Great work!</p>",
      "votes": null,
      "replies": [
        {
          "id": 907154,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "06/29/2020 18:02:04",
          "content": "<p>Thanks <a href=\"/anmspro\">@anmspro</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 907204,
      "author_name": "yash612",
      "author_url": "",
      "post_date": "06/29/2020 18:56:58",
      "content": "<p>DETR is very interesting , I heard it's very good for segmentation too.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "906904": "Ever since their inception transformers, the main research has been around trying to leverage them in different tasks apart from NLP. Facebook research launched a recent paper END TO END OBJECT DETECTION USING TRANSFORMERS :DETR. It uses transformers for Detection of objects using a new loss function. I have written a pipeline to fine-tune DETR on custom dataset using this competition to achieve state of the art results. \n\nhttps://www.kaggle.com/tanulsingh077/end-to-end-object-detection-with-transformers-detr\n\nDETR gives a fairly good result with training on a single epoch itself , being a transformer model , it does not require a lot of epoch to train, also with lr strategy. I am planning to update this kernel after complete training and also more kernels on this topic.",
    "906924": "Great notebook! I would like to ask you one question, can you tell how good it performs as compared to YOLOv5?",
    "907040": "Great work!",
    "907124": "I am still in early stages and I have not fully started with fine tuning yet, I will start tomorrow after m5 , maybe then I will update results here",
    "907152": "Cool! Waiting to see the benchmarks :)",
    "907154": "Thanks @anmspro",
    "907204": "DETR is very interesting , I heard it's very good for segmentation too.",
    "963268": "Great piece of work! I have tried to re implement the DeTr with some changement in the minimal code they have shared in the original paper. Unfortunatly my re implementation do not work well, I think I know where is located the issue since I am not using the NestedTensor (I think they do the absraction for no object labels )in my architecture. Do someone have tried to do the same experience?"
  },
  "source": "meta"
}