{
  "id": 54599,
  "title": "Our solution sketch: public LB 0.512",
  "url": "/competitions/data-science-bowl-2018/discussion/54599",
  "author_name": "",
  "post_date": "2018-04-15T17:15:22.911741800Z",
  "votes": 25,
  "comment_count": 4,
  "views": 0,
  "content": "<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/314490/9193/dsb.png\" alt=\"DSB solution\">\nWhile we are waiting for the inference to run, here is a rough sketch of our solution:</p>\n\n<ol>\n<li>RetinaNet for RPN</li>\n<li>Crop and resize directly based on the image using RPN predictions</li>\n<li>VGG 16 based classifier</li>\n<li>U-Net based segmentation </li>\n</ol>\n\n<p>The above got close 0.48 on the LB with some tuning. With TTA, we were able to get to about 0.51. TTA included 5 scaled images for RPN, 8 rotations of each crop for classifier and segmentation predictions.</p>\n\n<p>Data augmentation technique ended up being rotations/flips though not a major difference in performance.</p>\n\n<p>All the three models were trained separately. This way, the training  ended up being a lot more stable for us than training end-to-end mask R-CNN.</p>\n\n<p>Thank you all for the great discussion, and thanks to Heng CherKeng interesting discussion, code and ideas!</p>\n\n<p>Thanks to the organizes for a great competition, and Allen in particular for the guidance.</p>",
  "messages": [
    {
      "id": "314490",
      "postDate": "04/15/2018 17:15:22",
      "content": "<p><img src=\"https://storage.googleapis.com/kaggle-forum-message-attachments/314490/9193/dsb.png\" alt=\"DSB solution\">\nWhile we are waiting for the inference to run, here is a rough sketch of our solution:</p>\n\n<ol>\n<li>RetinaNet for RPN</li>\n<li>Crop and resize directly based on the image using RPN predictions</li>\n<li>VGG 16 based classifier</li>\n<li>U-Net based segmentation </li>\n</ol>\n\n<p>The above got close 0.48 on the LB with some tuning. With TTA, we were able to get to about 0.51. TTA included 5 scaled images for RPN, 8 rotations of each crop for classifier and segmentation predictions.</p>\n\n<p>Data augmentation technique ended up being rotations/flips though not a major difference in performance.</p>\n\n<p>All the three models were trained separately. This way, the training  ended up being a lot more stable for us than training end-to-end mask R-CNN.</p>\n\n<p>Thank you all for the great discussion, and thanks to Heng CherKeng interesting discussion, code and ideas!</p>\n\n<p>Thanks to the organizes for a great competition, and Allen in particular for the guidance.</p>",
      "rawMarkdown": "![DSB solution] (https://kaggle2.blob.core.windows.net/forum-message-attachments/314490/9193/dsb.png)\nWhile we are waiting for the inference to run, here is a rough sketch of our solution:\n\n 1. RetinaNet for RPN\n 2. Crop and resize directly based on the image using RPN predictions\n 3. VGG 16 based classifier\n 4. U-Net based segmentation \n\nThe above got close 0.48 on the LB with some tuning. With TTA, we were able to get to about 0.51. TTA included 5 scaled images for RPN, 8 rotations of each crop for classifier and segmentation predictions.\n\nData augmentation technique ended up being rotations/flips though not a major difference in performance.\n\nAll the three models were trained separately. This way, the training  ended up being a lot more stable for us than training end-to-end mask R-CNN.\n\nThank you all for the great discussion, and thanks to Heng CherKeng interesting discussion, code and ideas!\n\nThanks to the organizes for a great competition, and Allen in particular for the guidance.",
      "votes": null
    },
    {
      "id": "314734",
      "postDate": "04/16/2018 07:18:42",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": null
    },
    {
      "id": "315066",
      "postDate": "04/16/2018 17:54:21",
      "content": "<p>I started with retina net but I saw he didn't detected all boxes and it took him a lot to run. </p>",
      "rawMarkdown": "I started with retina net but I saw he didn't detected all boxes and it took him a lot to run.",
      "votes": null
    },
    {
      "id": "315173",
      "postDate": "04/16/2018 20:19:17",
      "content": "<p>RetinaNet worked ok for us though it did end up being the bottleneck of the performance. We did have to change the default implementation of using P3 to P7 of feature pyramids to use P1 to P5 along with lower strides.  Training was slow but not too bad. </p>",
      "rawMarkdown": "RetinaNet worked ok for us though it did end up being the bottleneck of the performance. We did have to change the default implementation of using P3 to P7 of feature pyramids to use P1 to P5 along with lower strides.  Training was slow but not too bad.",
      "votes": null
    },
    {
      "id": "315275",
      "postDate": "04/17/2018 00:35:29",
      "content": "<p>Congratulations to all the winners! </p>\n\n<p>Private LB of of the above solution was 0.571. We had a submission w/o retraining based on stage1 solution that achieved 0.585, but we didnt select it.  </p>",
      "rawMarkdown": "Congratulations to all the winners! \n\nPrivate LB of of the above solution was 0.571. We had a submission w/o retraining based on stage1 solution that achieved 0.585, but we didnt select it.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 314734,
      "author_name": "florianhuber",
      "author_url": "",
      "post_date": "04/16/2018 07:18:42",
      "content": "<p>Thanks for sharing!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 315066,
      "author_name": "aviwolfson",
      "author_url": "",
      "post_date": "04/16/2018 17:54:21",
      "content": "<p>I started with retina net but I saw he didn't detected all boxes and it took him a lot to run. </p>",
      "votes": null,
      "replies": [
        {
          "id": 315173,
          "author_name": "yuanqing",
          "author_url": "",
          "post_date": "04/16/2018 20:19:17",
          "content": "<p>RetinaNet worked ok for us though it did end up being the bottleneck of the performance. We did have to change the default implementation of using P3 to P7 of feature pyramids to use P1 to P5 along with lower strides.  Training was slow but not too bad. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 315275,
      "author_name": "yuanqing",
      "author_url": "",
      "post_date": "04/17/2018 00:35:29",
      "content": "<p>Congratulations to all the winners! </p>\n\n<p>Private LB of of the above solution was 0.571. We had a submission w/o retraining based on stage1 solution that achieved 0.585, but we didnt select it.  </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "314490": "![DSB solution] (https://kaggle2.blob.core.windows.net/forum-message-attachments/314490/9193/dsb.png)\nWhile we are waiting for the inference to run, here is a rough sketch of our solution:\n\n 1. RetinaNet for RPN\n 2. Crop and resize directly based on the image using RPN predictions\n 3. VGG 16 based classifier\n 4. U-Net based segmentation \n\nThe above got close 0.48 on the LB with some tuning. With TTA, we were able to get to about 0.51. TTA included 5 scaled images for RPN, 8 rotations of each crop for classifier and segmentation predictions.\n\nData augmentation technique ended up being rotations/flips though not a major difference in performance.\n\nAll the three models were trained separately. This way, the training  ended up being a lot more stable for us than training end-to-end mask R-CNN.\n\nThank you all for the great discussion, and thanks to Heng CherKeng interesting discussion, code and ideas!\n\nThanks to the organizes for a great competition, and Allen in particular for the guidance.",
    "314734": "Thanks for sharing!",
    "315066": "I started with retina net but I saw he didn't detected all boxes and it took him a lot to run.",
    "315173": "RetinaNet worked ok for us though it did end up being the bottleneck of the performance. We did have to change the default implementation of using P3 to P7 of feature pyramids to use P1 to P5 along with lower strides.  Training was slow but not too bad.",
    "315275": "Congratulations to all the winners! \n\nPrivate LB of of the above solution was 0.571. We had a submission w/o retraining based on stage1 solution that achieved 0.585, but we didnt select it."
  },
  "source": "meta"
}