{
  "id": 561414,
  "title": "35th place solution [2D + 3D + YOLO]",
  "url": "/competitions/czii-cryo-et-object-identification/discussion/561414",
  "author_name": "",
  "post_date": "2025-02-06T02:15:51.811745700Z",
  "votes": 12,
  "comment_count": 2,
  "views": 0,
  "content": "<p>This is my first silver medal with kaggle. I truly appreciate this opportunity to experiment interesting methods.</p>\n<p><strong>Datasets</strong><br>\nOriginal training dataset (DS-10440) + Extra (DS-10441) with gaussian blur</p>\n<p><strong>Models</strong><br>\n2D : Segformer with Mixed Vision Transformer B2 Encoder<br>\n3D : SegResNet with Deep Supervision<br>\nYOLO : YOLOv8</p>\n<p><strong>Training</strong><br>\n2D : EMA decay 0.99<br>\n3D : EMA decay 0.99 + 36 stride patch training + 72 stride patch validation<br>\nYOLO : Nothing special but compared to v11 it shows higher recall [<a href=\"https://github.com/ultralytics/ultralytics/issues/16966\" target=\"_blank\">https://github.com/ultralytics/ultralytics/issues/16966</a>]</p>\n<p><strong>Final Validation</strong><br>\n2D + 3D + YOLO 7 Fold<br>\n0.25 2D + 0.75 3D logits ensemble + YOLO (apo-ferretin, ribosome, vlp)<br>\nDBScan with 0.8*radius</p>\n<p>Code would be updated soon. Thanks.</p>",
  "messages": [
    {
      "id": "3116476",
      "postDate": "02/06/2025 02:15:51",
      "content": "<p>This is my first silver medal with kaggle. I truly appreciate this opportunity to experiment interesting methods.</p>\n<p><strong>Datasets</strong><br>\nOriginal training dataset (DS-10440) + Extra (DS-10441) with gaussian blur</p>\n<p><strong>Models</strong><br>\n2D : Segformer with Mixed Vision Transformer B2 Encoder<br>\n3D : SegResNet with Deep Supervision<br>\nYOLO : YOLOv8</p>\n<p><strong>Training</strong><br>\n2D : EMA decay 0.99<br>\n3D : EMA decay 0.99 + 36 stride patch training + 72 stride patch validation<br>\nYOLO : Nothing special but compared to v11 it shows higher recall [<a href=\"https://github.com/ultralytics/ultralytics/issues/16966\" target=\"_blank\">https://github.com/ultralytics/ultralytics/issues/16966</a>]</p>\n<p><strong>Final Validation</strong><br>\n2D + 3D + YOLO 7 Fold<br>\n0.25 2D + 0.75 3D logits ensemble + YOLO (apo-ferretin, ribosome, vlp)<br>\nDBScan with 0.8*radius</p>\n<p>Code would be updated soon. Thanks.</p>",
      "rawMarkdown": "This is my first silver medal with kaggle. I truly appreciate this opportunity to experiment interesting methods.\n\n**Datasets**\nOriginal training dataset (DS-10440) + Extra (DS-10441) with gaussian blur\n\n**Models**\n2D : Segformer with Mixed Vision Transformer B2 Encoder\n3D : SegResNet with Deep Supervision\nYOLO : YOLOv8\n\n**Training**\n2D : EMA decay 0.99\n3D : EMA decay 0.99 + 36 stride patch training + 72 stride patch validation\nYOLO : Nothing special but compared to v11 it shows higher recall [https://github.com/ultralytics/ultralytics/issues/16966]\n\n**Final Validation**\n2D + 3D + YOLO 7 Fold\n0.25 2D + 0.75 3D logits ensemble + YOLO (apo-ferretin, ribosome, vlp)\nDBScan with 0.8*radius\n\n\nCode would be updated soon. Thanks.",
      "votes": null
    },
    {
      "id": "3116745",
      "postDate": "02/06/2025 09:09:21",
      "content": "<p>Congratulations! Are you planning to share your code as well? I would love to read through it.</p>",
      "rawMarkdown": "Congratulations! Are you planning to share your code as well? I would love to read through it.",
      "votes": null
    },
    {
      "id": "3116749",
      "postDate": "02/06/2025 09:11:34",
      "content": "<p>Thank you :) and sure. It would be uploaded soon!</p>",
      "rawMarkdown": "Thank you :) and sure. It would be uploaded soon!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 3116745,
      "author_name": "snnclsr",
      "author_url": "",
      "post_date": "02/06/2025 09:09:21",
      "content": "<p>Congratulations! Are you planning to share your code as well? I would love to read through it.</p>",
      "votes": null,
      "replies": [
        {
          "id": 3116749,
          "author_name": "jungchulkoo",
          "author_url": "",
          "post_date": "02/06/2025 09:11:34",
          "content": "<p>Thank you :) and sure. It would be uploaded soon!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "3116476": "This is my first silver medal with kaggle. I truly appreciate this opportunity to experiment interesting methods.\n\n**Datasets**\nOriginal training dataset (DS-10440) + Extra (DS-10441) with gaussian blur\n\n**Models**\n2D : Segformer with Mixed Vision Transformer B2 Encoder\n3D : SegResNet with Deep Supervision\nYOLO : YOLOv8\n\n**Training**\n2D : EMA decay 0.99\n3D : EMA decay 0.99 + 36 stride patch training + 72 stride patch validation\nYOLO : Nothing special but compared to v11 it shows higher recall [https://github.com/ultralytics/ultralytics/issues/16966]\n\n**Final Validation**\n2D + 3D + YOLO 7 Fold\n0.25 2D + 0.75 3D logits ensemble + YOLO (apo-ferretin, ribosome, vlp)\nDBScan with 0.8*radius\n\n\nCode would be updated soon. Thanks.",
    "3116745": "Congratulations! Are you planning to share your code as well? I would love to read through it.",
    "3116749": "Thank you :) and sure. It would be uploaded soon!"
  },
  "source": "meta"
}