{
  "id": 402866,
  "title": "Some ideas for exploration",
  "url": "/competitions/vesuvius-challenge-ink-detection/discussion/402866",
  "author_name": "",
  "post_date": "2023-04-20T02:30:16.682792700Z",
  "votes": -4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Just throwing out some ideas related to the present competition.</p>\n<p>Deep learning. This appears to be an active arena of investigation of the problem at hand. 3D convolutional neural network architectures like 3D-UNets or 3D ResNets used to capture and reveal spatial dependencies in the data. Under this heading, I would also include transformer-based models adapted to 3D data, which have been comparatively successful in computer vision. I don't know if anyone is doing that here, but it seems like promising path.</p>\n<p>What about leveraging models from related fields? Medical imaging, for example, or materials science. Would it be possible to use a pre-trained model from a related domain and fine tune it to the ink detection problem? This would, perhaps, accelerate the development of a robust solution. </p>\n<p>Given the paucity of training examples, coherent means of data augmentation might be a useful avenue to explore.</p>\n<p>And then, of course, there are the seemingly endless ensembles of models. </p>\n<p>It also seems like active learning might be useful in the present context.</p>",
  "messages": [
    {
      "id": "2227763",
      "postDate": "04/20/2023 02:30:16",
      "content": "<p>Just throwing out some ideas related to the present competition.</p>\n<p>Deep learning. This appears to be an active arena of investigation of the problem at hand. 3D convolutional neural network architectures like 3D-UNets or 3D ResNets used to capture and reveal spatial dependencies in the data. Under this heading, I would also include transformer-based models adapted to 3D data, which have been comparatively successful in computer vision. I don't know if anyone is doing that here, but it seems like promising path.</p>\n<p>What about leveraging models from related fields? Medical imaging, for example, or materials science. Would it be possible to use a pre-trained model from a related domain and fine tune it to the ink detection problem? This would, perhaps, accelerate the development of a robust solution. </p>\n<p>Given the paucity of training examples, coherent means of data augmentation might be a useful avenue to explore.</p>\n<p>And then, of course, there are the seemingly endless ensembles of models. </p>\n<p>It also seems like active learning might be useful in the present context.</p>",
      "rawMarkdown": "Just throwing out some ideas related to the present competition.\n\nDeep learning. This appears to be an active arena of investigation of the problem at hand. 3D convolutional neural network architectures like 3D-UNets or 3D ResNets used to capture and reveal spatial dependencies in the data. Under this heading, I would also include transformer-based models adapted to 3D data, which have been comparatively successful in computer vision. I don't know if anyone is doing that here, but it seems like promising path.\n\nWhat about leveraging models from related fields? Medical imaging, for example, or materials science. Would it be possible to use a pre-trained model from a related domain and fine tune it to the ink detection problem? This would, perhaps, accelerate the development of a robust solution. \n\nGiven the paucity of training examples, coherent means of data augmentation might be a useful avenue to explore.\n\nAnd then, of course, there are the seemingly endless ensembles of models. \n\nIt also seems like active learning might be useful in the present context.",
      "votes": null
    },
    {
      "id": "2228068",
      "postDate": "04/20/2023 08:48:06",
      "content": "<p>Yes, have a look on public notebooks, your ideas are already there !</p>",
      "rawMarkdown": "Yes, have a look on public notebooks, your ideas are already there !",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2228068,
      "author_name": "iraqbot",
      "author_url": "",
      "post_date": "04/20/2023 08:48:06",
      "content": "<p>Yes, have a look on public notebooks, your ideas are already there !</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2227763": "Just throwing out some ideas related to the present competition.\n\nDeep learning. This appears to be an active arena of investigation of the problem at hand. 3D convolutional neural network architectures like 3D-UNets or 3D ResNets used to capture and reveal spatial dependencies in the data. Under this heading, I would also include transformer-based models adapted to 3D data, which have been comparatively successful in computer vision. I don't know if anyone is doing that here, but it seems like promising path.\n\nWhat about leveraging models from related fields? Medical imaging, for example, or materials science. Would it be possible to use a pre-trained model from a related domain and fine tune it to the ink detection problem? This would, perhaps, accelerate the development of a robust solution. \n\nGiven the paucity of training examples, coherent means of data augmentation might be a useful avenue to explore.\n\nAnd then, of course, there are the seemingly endless ensembles of models. \n\nIt also seems like active learning might be useful in the present context.",
    "2228068": "Yes, have a look on public notebooks, your ideas are already there !"
  },
  "source": "meta"
}