{
  "id": 278951,
  "title": "common tricks in kaggle image segmentation problems ",
  "url": "/competitions/sartorius-cell-instance-segmentation/discussion/278951",
  "author_name": "",
  "post_date": "2021-10-16T05:17:36.467334800Z",
  "votes": 20,
  "comment_count": 2,
  "views": 0,
  "content": "<p>this thread will be updated as I make progress</p>\n<ol>\n<li>clustering - cluster similar patches together. KNN neighbours can be used in pre-post processing, data augmentation/balancing etc. clustering can be done for input or output or paired input-output</li>\n<li>Train a classifier to predict kaggle metric score (LB score). This can be used for post-processing.</li>\n<li>combine the patches into a bigger image (jigsaw puzzle problem).</li>\n<li>pseudo-labeling, semi-supervised learning, knowledge distillation, adversarial training</li>\n<li>more labeling or supervisory signal (regularization by additional learning task). e.g create distance transform regression target and train network to predict it.</li>\n<li>train binary classifier to predict if the mask is empty or not. Train another network for segmentation.</li>\n<li>instead of single class segmentation, you can consider multiclass (e.g. treat mask of different size, location, different image texture as a different class)</li>\n<li>label train/sample as difficult or easy using prediction confidence. break big problems into smaller ones. difficult samples can have more post/pre-processing, etc … </li>\n</ol>",
  "messages": [
    {
      "id": "1546381",
      "postDate": "10/16/2021 05:17:36",
      "content": "<p>this thread will be updated as I make progress</p>\n<ol>\n<li>clustering - cluster similar patches together. KNN neighbours can be used in pre-post processing, data augmentation/balancing etc. clustering can be done for input or output or paired input-output</li>\n<li>Train a classifier to predict kaggle metric score (LB score). This can be used for post-processing.</li>\n<li>combine the patches into a bigger image (jigsaw puzzle problem).</li>\n<li>pseudo-labeling, semi-supervised learning, knowledge distillation, adversarial training</li>\n<li>more labeling or supervisory signal (regularization by additional learning task). e.g create distance transform regression target and train network to predict it.</li>\n<li>train binary classifier to predict if the mask is empty or not. Train another network for segmentation.</li>\n<li>instead of single class segmentation, you can consider multiclass (e.g. treat mask of different size, location, different image texture as a different class)</li>\n<li>label train/sample as difficult or easy using prediction confidence. break big problems into smaller ones. difficult samples can have more post/pre-processing, etc … </li>\n</ol>",
      "rawMarkdown": "this thread will be updated as I make progress\n\n1. clustering - cluster similar patches together. KNN neighbours can be used in pre-post processing, data augmentation/balancing etc. clustering can be done for input or output or paired input-output\n2. Train a classifier to predict kaggle metric score (LB score). This can be used for post-processing.\n3. combine the patches into a bigger image (jigsaw puzzle problem).\n4. pseudo-labeling, semi-supervised learning, knowledge distillation, adversarial training\n5. more labeling or supervisory signal (regularization by additional learning task). e.g create distance transform regression target and train network to predict it.\n6. train binary classifier to predict if the mask is empty or not. Train another network for segmentation.\n7. instead of single class segmentation, you can consider multiclass (e.g. treat mask of different size, location, different image texture as a different class)\n8. label train/sample as difficult or easy using prediction confidence. break big problems into smaller ones. difficult samples can have more post/pre-processing, etc …",
      "votes": null
    },
    {
      "id": "1546810",
      "postDate": "10/16/2021 14:32:56",
      "content": "<p>thanks mahdi very useful.</p>",
      "rawMarkdown": "thanks mahdi very useful.",
      "votes": null
    },
    {
      "id": "1623387",
      "postDate": "12/19/2021 20:06:20",
      "content": "<p>Thank you for sharing. My feeling is some of the tricks may decrease the robustness of our models.</p>",
      "rawMarkdown": "Thank you for sharing. My feeling is some of the tricks may decrease the robustness of our models.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1546810,
      "author_name": "hadisahmadian",
      "author_url": "",
      "post_date": "10/16/2021 14:32:56",
      "content": "<p>thanks mahdi very useful.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1623387,
      "author_name": "matrixoutsider",
      "author_url": "",
      "post_date": "12/19/2021 20:06:20",
      "content": "<p>Thank you for sharing. My feeling is some of the tricks may decrease the robustness of our models.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1546381": "this thread will be updated as I make progress\n\n1. clustering - cluster similar patches together. KNN neighbours can be used in pre-post processing, data augmentation/balancing etc. clustering can be done for input or output or paired input-output\n2. Train a classifier to predict kaggle metric score (LB score). This can be used for post-processing.\n3. combine the patches into a bigger image (jigsaw puzzle problem).\n4. pseudo-labeling, semi-supervised learning, knowledge distillation, adversarial training\n5. more labeling or supervisory signal (regularization by additional learning task). e.g create distance transform regression target and train network to predict it.\n6. train binary classifier to predict if the mask is empty or not. Train another network for segmentation.\n7. instead of single class segmentation, you can consider multiclass (e.g. treat mask of different size, location, different image texture as a different class)\n8. label train/sample as difficult or easy using prediction confidence. break big problems into smaller ones. difficult samples can have more post/pre-processing, etc …",
    "1546810": "thanks mahdi very useful.",
    "1623387": "Thank you for sharing. My feeling is some of the tricks may decrease the robustness of our models."
  },
  "source": "meta"
}