{
  "id": 208542,
  "title": "LB 0.881 | SE-ResNeXT + EfficientNet-B5 with Depth Estimation Network",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/208542",
  "author_name": "lukachkhetiani",
  "post_date": "2021-01-03T22:17:35.611000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Image masking with Monocular Depth Estimation Network gave ~7% boost over baseline. <br>\nBasically we take depth map and filter everything out via (min + max)/3. This way, we don't have to handle very noisy backgrounds. <br>\nIt was very much uncomfortable to train such deep models for this particular task, as long as features aren't really difficult, but ensemble + depth map filtering is worth trying. </p>\n<p>See the notebook: <br>\n<a href=\"https://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble\" target=\"_blank\">https://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble</a></p>",
  "messages": [
    {
      "id": 1137394,
      "postDate": "2021-01-03T22:17:35.610Z",
      "content": "<p>Image masking with Monocular Depth Estimation Network gave ~7% boost over baseline. <br>\nBasically we take depth map and filter everything out via (min + max)/3. This way, we don't have to handle very noisy backgrounds. <br>\nIt was very much uncomfortable to train such deep models for this particular task, as long as features aren't really difficult, but ensemble + depth map filtering is worth trying. </p>\n<p>See the notebook: <br>\n<a href=\"https://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble\" target=\"_blank\">https://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble</a></p>",
      "rawMarkdown": "Image masking with Monocular Depth Estimation Network gave ~7% boost over baseline. \nBasically we take depth map and filter everything out via (min + max)/3. This way, we don't have to handle very noisy backgrounds. \nIt was very much uncomfortable to train such deep models for this particular task, as long as features aren't really difficult, but ensemble + depth map filtering is worth trying. \n\nSee the notebook: \nhttps://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1137394": "Image masking with Monocular Depth Estimation Network gave ~7% boost over baseline. \nBasically we take depth map and filter everything out via (min + max)/3. This way, we don't have to handle very noisy backgrounds. \nIt was very much uncomfortable to train such deep models for this particular task, as long as features aren't really difficult, but ensemble + depth map filtering is worth trying. \n\nSee the notebook: \nhttps://www.kaggle.com/lukachkhetiani/se-resnext-101-efficientnet-b5-ensemble"
  }
}