{
  "id": 614057,
  "title": "Tips - Bounding Box embeddings",
  "url": "/competitions/the-3lc-cotton-weed-detection-challenge/discussion/614057",
  "author_name": "PaulEE",
  "post_date": "2025-10-31T17:49:13.610000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Extending the tables with embeddings per Bounding Box can give good insights! </p>\n<p>clone <a href=\"https://github.com/3lc-ai/3lc-examples\" target=\"_blank\">https://github.com/3lc-ai/3lc-examples</a>\npip install -e .\nthen run a script that trains a classifier of whats inside the tables BB's like this:</p>\n<p>3lc-tools run augment-instance-table --train_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-train1 --val_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-val1 --epochs 12</p>\n<p>This creates two new revisions of the train and val set - extending the tables with per BB embedding - now it's easy to find where the model struggles, labels thats possible wrong etc - use lasso tool to select regions and QC / fix labels!</p>\n<p>Are these labels to the right correct?</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262906%2F49fd1650df97be61ca7d76526fa8f7e6%2FKaggleBBEmbeddings.jpg?generation=1761932859740011&amp;alt=media\" alt=\"\"></p>",
  "messages": [
    {
      "id": 3309449,
      "postDate": "2025-10-31T17:49:13.610Z",
      "content": "<p>Extending the tables with embeddings per Bounding Box can give good insights! </p>\n<p>clone <a href=\"https://github.com/3lc-ai/3lc-examples\" target=\"_blank\">https://github.com/3lc-ai/3lc-examples</a>\npip install -e .\nthen run a script that trains a classifier of whats inside the tables BB's like this:</p>\n<p>3lc-tools run augment-instance-table --train_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-train1 --val_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-val1 --epochs 12</p>\n<p>This creates two new revisions of the train and val set - extending the tables with per BB embedding - now it's easy to find where the model struggles, labels thats possible wrong etc - use lasso tool to select regions and QC / fix labels!</p>\n<p>Are these labels to the right correct?</p>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262906%2F49fd1650df97be61ca7d76526fa8f7e6%2FKaggleBBEmbeddings.jpg?generation=1761932859740011&amp;alt=media\" alt=\"\"></p>",
      "rawMarkdown": "Extending the tables with embeddings per Bounding Box can give good insights! \n\nclone https://github.com/3lc-ai/3lc-examples\npip install -e .\nthen run a script that trains a classifier of whats inside the tables BB's like this:\n\n3lc-tools run augment-instance-table --train_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-train1 --val_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-val1 --epochs 12\n\nThis creates two new revisions of the train and val set - extending the tables with per BB embedding - now it's easy to find where the model struggles, labels thats possible wrong etc - use lasso tool to select regions and QC / fix labels!\n\nAre these labels to the right correct?\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262906%2F49fd1650df97be61ca7d76526fa8f7e6%2FKaggleBBEmbeddings.jpg?generation=1761932859740011&alt=media)"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3309449": "Extending the tables with embeddings per Bounding Box can give good insights! \n\nclone https://github.com/3lc-ai/3lc-examples\npip install -e .\nthen run a script that trains a classifier of whats inside the tables BB's like this:\n\n3lc-tools run augment-instance-table --train_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-train1 --val_table /home/ubuntu/3LC/kaggle/datasets/cotton_weed_det3/tables/cotton_weed_det3-val1 --epochs 12\n\nThis creates two new revisions of the train and val set - extending the tables with per BB embedding - now it's easy to find where the model struggles, labels thats possible wrong etc - use lasso tool to select regions and QC / fix labels!\n\nAre these labels to the right correct?\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262906%2F49fd1650df97be61ca7d76526fa8f7e6%2FKaggleBBEmbeddings.jpg?generation=1761932859740011&alt=media)"
  }
}