{
  "id": 311941,
  "title": "Simple Ideas",
  "url": "/competitions/ultra-mnist/discussion/311941",
  "author_name": "Awsaf",
  "post_date": "2022-03-09T15:42:43.046000",
  "votes": 26,
  "comment_count": 18,
  "views": 0,
  "content": "<p>Here are some simple ideas to solve this problem,</p>\n<h2>Object Detection</h2>\n<ul>\n<li>Use Object Detection Model to generate bound box for each digit and simply add them</li>\n<li>For bounding box either you can label the training dataset (which I believe allowed) or you can look for them on the internet, Here' a similar dataset <a href=\"https://github.com/hukkelas/MNIST-ObjectDetection\" target=\"_blank\">https://github.com/hukkelas/MNIST-ObjectDetection</a><br>\n<img src=\"https://i.ibb.co/p0DVRLD/example.png\" alt=\"example\"></li>\n<li>Just reminder dataset has a check-board pattern(white-black) so existing model may not work quite well here.</li>\n</ul>\n<h2>Multi-Class Classification</h2>\n<ul>\n<li>As we know there will be images with sum [0-27] then we can simply do a 28 class classification.</li>\n<li>Looking at the image I'm not sure if the classification task will be easy. As image will vary a lot within a class.</li>\n</ul>",
  "messages": [
    {
      "id": 1717015,
      "postDate": "2022-03-09T15:42:43.047Z",
      "content": "<p>Here are some simple ideas to solve this problem,</p>\n<h2>Object Detection</h2>\n<ul>\n<li>Use Object Detection Model to generate bound box for each digit and simply add them</li>\n<li>For bounding box either you can label the training dataset (which I believe allowed) or you can look for them on the internet, Here' a similar dataset <a href=\"https://github.com/hukkelas/MNIST-ObjectDetection\" target=\"_blank\">https://github.com/hukkelas/MNIST-ObjectDetection</a><br>\n<img src=\"https://i.ibb.co/p0DVRLD/example.png\" alt=\"example\"></li>\n<li>Just reminder dataset has a check-board pattern(white-black) so existing model may not work quite well here.</li>\n</ul>\n<h2>Multi-Class Classification</h2>\n<ul>\n<li>As we know there will be images with sum [0-27] then we can simply do a 28 class classification.</li>\n<li>Looking at the image I'm not sure if the classification task will be easy. As image will vary a lot within a class.</li>\n</ul>",
      "rawMarkdown": "Here are some simple ideas to solve this problem,\n## Object Detection\n* Use Object Detection Model to generate bound box for each digit and simply add them\n* For bounding box either you can label the training dataset (which I believe allowed) or you can look for them on the internet, Here' a similar dataset https://github.com/hukkelas/MNIST-ObjectDetection\n<img src=\"https://i.ibb.co/p0DVRLD/example.png\" alt=\"example\" border=\"0\">\n* Just reminder dataset has a check-board pattern(white-black) so existing model may not work quite well here.\n\n## Multi-Class Classification\n* As we know there will be images with sum [0-27] then we can simply do a 28 class classification.\n* Looking at the image I'm not sure if the classification task will be easy. As image will vary a lot within a class.\n",
      "votes": 26
    },
    {
      "id": 1717027,
      "postDate": "2022-03-09T15:56:03.760Z",
      "content": "<p>Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.</p>",
      "rawMarkdown": "Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.",
      "votes": 6,
      "replies": [
        {
          "id": 1717059,
          "postDate": "2022-03-09T16:22:04.763Z",
          "content": "<p>Can we use the normal mnist dataset ?</p>",
          "rawMarkdown": "Can we use the normal mnist dataset ?",
          "votes": 1
        },
        {
          "id": 1717074,
          "postDate": "2022-03-09T16:33:40.420Z",
          "content": "<p>No any sort of external data is not allowed to be eligible for prize.</p>",
          "rawMarkdown": "No any sort of external data is not allowed to be eligible for prize.",
          "votes": 1
        },
        {
          "id": 1717139,
          "postDate": "2022-03-09T17:21:54.647Z",
          "content": "<p>that means we can use pretrained weights on MNIST data?  because we cannot use MNIST data for sure.  </p>",
          "rawMarkdown": "that means we can use pretrained weights on MNIST data?  because we cannot use MNIST data for sure.  ",
          "votes": 1
        },
        {
          "id": 1717155,
          "postDate": "2022-03-09T17:32:46.657Z",
          "rawMarkdown": "",
          "votes": 3,
          "isDeleted": true
        },
        {
          "id": 1717291,
          "postDate": "2022-03-09T19:14:06.817Z",
          "content": "<p>Just to further clarify. If I train a model on some version of public MNIST data… and then make that model publicly available. And then use that model to pseudo label the dataset. This is allowed?</p>",
          "rawMarkdown": "Just to further clarify. If I train a model on some version of public MNIST data... and then make that model publicly available. And then use that model to pseudo label the dataset. This is allowed?",
          "votes": 2
        },
        {
          "id": 1717301,
          "postDate": "2022-03-09T19:21:04.593Z",
          "content": "<p>We will never be able to verify such models. So, it would be better if we stick to traditional imagenet pretrained models for example the ones available in <code>timm</code>.</p>",
          "rawMarkdown": "We will never be able to verify such models. So, it would be better if we stick to traditional imagenet pretrained models for example the ones available in `timm`."
        },
        {
          "id": 1717305,
          "postDate": "2022-03-09T19:26:15.617Z",
          "content": "<p>Excellent. Sounds good.</p>",
          "rawMarkdown": "Excellent. Sounds good."
        },
        {
          "id": 1717733,
          "postDate": "2022-03-10T07:04:38.510Z",
          "content": "<p>If we put our pretrained weights in a public dataset will that be ok ? we could also make a thread for that </p>",
          "rawMarkdown": "If we put our pretrained weights in a public dataset will that be ok ? we could also make a thread for that "
        },
        {
          "id": 1728723,
          "postDate": "2022-03-19T07:44:33.880Z",
          "content": "<p>I think No <a href=\"https://www.kaggle.com/mithilsalunkhe\" target=\"_blank\">@mithilsalunkhe</a> </p>",
          "rawMarkdown": "I think No @mithilsalunkhe "
        },
        {
          "id": 1730290,
          "postDate": "2022-03-21T05:10:35.673Z",
          "content": "<p><a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a> The pretrained model just needs to be public , is that the only criteria?</p>",
          "rawMarkdown": "@abhishek The pretrained model just needs to be public , is that the only criteria?"
        },
        {
          "id": 1731121,
          "postDate": "2022-03-22T02:14:20.440Z",
          "content": "<blockquote>\n  <p>Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.</p>\n</blockquote>\n<p>Hi <a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a>, these are old guidelines right. Can we use custom dataset according to new guidelines? Need confirmation.</p>",
          "rawMarkdown": "> Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.\n\nHi @abhishek, these are old guidelines right. Can we use custom dataset according to new guidelines? Need confirmation."
        }
      ]
    },
    {
      "id": 1717449,
      "postDate": "2022-03-10T00:01:39.957Z",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> my friend we meet each other on all CV competitions :) Nice to see you again - look how could we simplify processing pipeline: <a href=\"https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required\" target=\"_blank\">https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required</a> </p>",
      "rawMarkdown": "@awsaf49 my friend we meet each other on all CV competitions :) Nice to see you again - look how could we simplify processing pipeline: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required ",
      "votes": 1
    },
    {
      "id": 1717033,
      "postDate": "2022-03-09T16:02:58.833Z",
      "content": "<p>What about regression? </p>",
      "rawMarkdown": "What about regression? ",
      "votes": 1,
      "replies": [
        {
          "id": 1717141,
          "postDate": "2022-03-09T17:22:32.257Z",
          "content": "<p>can you ellaborate ?  like simply regressing every image with sum?</p>",
          "rawMarkdown": "can you ellaborate ?  like simply regressing every image with sum?"
        }
      ]
    },
    {
      "id": 1717103,
      "postDate": "2022-03-09T16:54:40.990Z",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> I am sure it will be …………………. fun …. and detector out of the box will have challange … :)<br>\n<a href=\"https://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see\" target=\"_blank\">https://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see</a></p>",
      "rawMarkdown": "@awsaf49 I am sure it will be ...................... fun .... and detector out of the box will have challange ... :)\nhttps://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see",
      "votes": 2
    },
    {
      "id": 1718659,
      "postDate": "2022-03-11T04:07:00.650Z",
      "content": "<p>Excellent Work</p>",
      "rawMarkdown": "Excellent Work"
    },
    {
      "id": 1717183,
      "postDate": "2022-03-09T17:52:24.683Z",
      "content": "<p>How about using canny edge detector and segment each image and do a prediction?</p>",
      "rawMarkdown": "How about using canny edge detector and segment each image and do a prediction?"
    }
  ],
  "comments": [
    {
      "id": 1717027,
      "author_name": "Abhishek Thakur",
      "author_url": "",
      "post_date": "2022-03-09T15:56:03.760000",
      "content": "<p>Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.</p>",
      "votes": 6,
      "replies": [
        {
          "id": 1717059,
          "author_name": "Mithil Salunkhe",
          "author_url": "",
          "post_date": "2022-03-09T16:22:04.763000",
          "content": "<p>Can we use the normal mnist dataset ?</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1717074,
          "author_name": "Udbhav Bamba",
          "author_url": "",
          "post_date": "2022-03-09T16:33:40.420000",
          "content": "<p>No any sort of external data is not allowed to be eligible for prize.</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1717139,
          "author_name": "Bibhabasu Mohapatra",
          "author_url": "",
          "post_date": "2022-03-09T17:21:54.647000",
          "content": "<p>that means we can use pretrained weights on MNIST data?  because we cannot use MNIST data for sure.  </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1717155,
          "author_name": "",
          "author_url": "",
          "post_date": "2022-03-09T17:32:46.657000",
          "content": "",
          "votes": 3,
          "replies": []
        },
        {
          "id": 1717291,
          "author_name": "Darien Schettler",
          "author_url": "",
          "post_date": "2022-03-09T19:14:06.817000",
          "content": "<p>Just to further clarify. If I train a model on some version of public MNIST data… and then make that model publicly available. And then use that model to pseudo label the dataset. This is allowed?</p>",
          "votes": 2,
          "replies": []
        },
        {
          "id": 1717301,
          "author_name": "Abhishek Thakur",
          "author_url": "",
          "post_date": "2022-03-09T19:21:04.593000",
          "content": "<p>We will never be able to verify such models. So, it would be better if we stick to traditional imagenet pretrained models for example the ones available in <code>timm</code>.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1717305,
          "author_name": "Darien Schettler",
          "author_url": "",
          "post_date": "2022-03-09T19:26:15.617000",
          "content": "<p>Excellent. Sounds good.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1717733,
          "author_name": "Mithil Salunkhe",
          "author_url": "",
          "post_date": "2022-03-10T07:04:38.510000",
          "content": "<p>If we put our pretrained weights in a public dataset will that be ok ? we could also make a thread for that </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1728723,
          "author_name": "Tan Phan",
          "author_url": "",
          "post_date": "2022-03-19T07:44:33.880000",
          "content": "<p>I think No <a href=\"https://www.kaggle.com/mithilsalunkhe\" target=\"_blank\">@mithilsalunkhe</a> </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1730290,
          "author_name": "Devansh Chowdhury",
          "author_url": "",
          "post_date": "2022-03-21T05:10:35.673000",
          "content": "<p><a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a> The pretrained model just needs to be public , is that the only criteria?</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1731121,
          "author_name": "Asapanna Rakesh",
          "author_url": "",
          "post_date": "2022-03-22T02:14:20.440000",
          "content": "<blockquote>\n  <p>Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.</p>\n</blockquote>\n<p>Hi <a href=\"https://www.kaggle.com/abhishek\" target=\"_blank\">@abhishek</a>, these are old guidelines right. Can we use custom dataset according to new guidelines? Need confirmation.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1717449,
      "author_name": "Remek Kinas",
      "author_url": "",
      "post_date": "2022-03-10T00:01:39.957000",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> my friend we meet each other on all CV competitions :) Nice to see you again - look how could we simplify processing pipeline: <a href=\"https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required\" target=\"_blank\">https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required</a> </p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 1717033,
      "author_name": "Vadim Irtlach",
      "author_url": "",
      "post_date": "2022-03-09T16:02:58.833000",
      "content": "<p>What about regression? </p>",
      "votes": 1,
      "replies": [
        {
          "id": 1717141,
          "author_name": "Bibhabasu Mohapatra",
          "author_url": "",
          "post_date": "2022-03-09T17:22:32.257000",
          "content": "<p>can you ellaborate ?  like simply regressing every image with sum?</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1717103,
      "author_name": "Remek Kinas",
      "author_url": "",
      "post_date": "2022-03-09T16:54:40.990000",
      "content": "<p><a href=\"https://www.kaggle.com/awsaf49\" target=\"_blank\">@awsaf49</a> I am sure it will be …………………. fun …. and detector out of the box will have challange … :)<br>\n<a href=\"https://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see\" target=\"_blank\">https://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see</a></p>",
      "votes": 2,
      "replies": []
    },
    {
      "id": 1718659,
      "author_name": "Shivam0105",
      "author_url": "",
      "post_date": "2022-03-11T04:07:00.650000",
      "content": "<p>Excellent Work</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1717183,
      "author_name": "Yashwanth",
      "author_url": "",
      "post_date": "2022-03-09T17:52:24.683000",
      "content": "<p>How about using canny edge detector and segment each image and do a prediction?</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1717015": "Here are some simple ideas to solve this problem,\n## Object Detection\n* Use Object Detection Model to generate bound box for each digit and simply add them\n* For bounding box either you can label the training dataset (which I believe allowed) or you can look for them on the internet, Here' a similar dataset https://github.com/hukkelas/MNIST-ObjectDetection\n<img src=\"https://i.ibb.co/p0DVRLD/example.png\" alt=\"example\" border=\"0\">\n* Just reminder dataset has a check-board pattern(white-black) so existing model may not work quite well here.\n\n## Multi-Class Classification\n* As we know there will be images with sum [0-27] then we can simply do a 28 class classification.\n* Looking at the image I'm not sure if the classification task will be easy. As image will vary a lot within a class.\n",
    "1717027": "Please keep in mind that external data is NOT allowed in this competition ;) You can, however, use pre-trained models which are publicly available to everyone.",
    "1717449": "@awsaf49 my friend we meet each other on all CV competitions :) Nice to see you again - look how could we simplify processing pipeline: https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required ",
    "1717033": "What about regression? ",
    "1717103": "@awsaf49 I am sure it will be ...................... fun .... and detector out of the box will have challange ... :)\nhttps://www.kaggle.com/remekkinas/funny-cv-eda-cases-we-see",
    "1718659": "Excellent Work",
    "1717183": "How about using canny edge detector and segment each image and do a prediction?"
  }
}