{
  "id": 319932,
  "title": "One tip - biased (squared) annotation to improve DS quality ",
  "url": "/competitions/happy-whale-and-dolphin/discussion/319932",
  "author_name": "",
  "post_date": "2022-04-19T12:54:36.323872400Z",
  "votes": 9,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Congratulations to everybody! Great reaults and really fun competition. A lot of tips to learn. <br>\nI am reading solution description and there is no new thing to add from my side. Outstanding solutions. One more tip from my side - consistent annotation to prepare data for specific input format. What does it mean?</p>\n<p>I noticed some score increase annotating dorsal fins such way they are not distorted after NN input resizing (to preserve shape of dorsal fin). All our dorsal fin were consciously annotated in form of square (w and h almost the same). The more consistent way (all dorsal fin annotated as close to square as possible) the better. As a result our dorsal fin detector was biased to square bboxes. Then when we put it to NN (before resizing to square eg. 360x360 or 768x768) we save characteristic shape of fin identity. </p>\n<p>Annotation - as close to square as possible<br>\n<img src=\"https://i.ibb.co/DWxTsFs/001.jpg\" alt=\"annot\"></p>\n<p>Result - no distorted dorsal fins<br>\n<img src=\"https://i.ibb.co/XZDjyZ2/002.jpg\" alt=\"result\"></p>\n<p>Another way was to put full body or dorsal fin using letterbox technique (commonly used in augumentation - eg. mosiac) . See below. Adding 0 pixels does not influence convolution. In this case (letterbox) more investigation is required to prove score increase.</p>\n<p><img src=\"https://i.ibb.co/Byf6sLs/003.jpg\" alt=\"letterbox\"></p>",
  "messages": [
    {
      "id": "1760661",
      "postDate": "04/19/2022 12:54:36",
      "content": "<p>Congratulations to everybody! Great reaults and really fun competition. A lot of tips to learn. <br>\nI am reading solution description and there is no new thing to add from my side. Outstanding solutions. One more tip from my side - consistent annotation to prepare data for specific input format. What does it mean?</p>\n<p>I noticed some score increase annotating dorsal fins such way they are not distorted after NN input resizing (to preserve shape of dorsal fin). All our dorsal fin were consciously annotated in form of square (w and h almost the same). The more consistent way (all dorsal fin annotated as close to square as possible) the better. As a result our dorsal fin detector was biased to square bboxes. Then when we put it to NN (before resizing to square eg. 360x360 or 768x768) we save characteristic shape of fin identity. </p>\n<p>Annotation - as close to square as possible<br>\n<img src=\"https://i.ibb.co/DWxTsFs/001.jpg\" alt=\"annot\"></p>\n<p>Result - no distorted dorsal fins<br>\n<img src=\"https://i.ibb.co/XZDjyZ2/002.jpg\" alt=\"result\"></p>\n<p>Another way was to put full body or dorsal fin using letterbox technique (commonly used in augumentation - eg. mosiac) . See below. Adding 0 pixels does not influence convolution. In this case (letterbox) more investigation is required to prove score increase.</p>\n<p><img src=\"https://i.ibb.co/Byf6sLs/003.jpg\" alt=\"letterbox\"></p>",
      "rawMarkdown": "Congratulations to everybody! Great reaults and really fun competition. A lot of tips to learn. \nI am reading solution description and there is no new thing to add from my side. Outstanding solutions. One more tip from my side - consistent annotation to prepare data for specific input format. What does it mean?\n\nI noticed some score increase annotating dorsal fins such way they are not distorted after NN input resizing (to preserve shape of dorsal fin). All our dorsal fin were consciously annotated in form of square (w and h almost the same). The more consistent way (all dorsal fin annotated as close to square as possible) the better. As a result our dorsal fin detector was biased to square bboxes. Then when we put it to NN (before resizing to square eg. 360x360 or 768x768) we save characteristic shape of fin identity. \n\nAnnotation - as close to square as possible\n![annot](https://i.ibb.co/DWxTsFs/001.jpg)\n\nResult - no distorted dorsal fins\n![result](https://i.ibb.co/XZDjyZ2/002.jpg)\n\nAnother way was to put full body or dorsal fin using letterbox technique (commonly used in augumentation - eg. mosiac) . See below. Adding 0 pixels does not influence convolution. In this case (letterbox) more investigation is required to prove score increase.\n\n![letterbox](https://i.ibb.co/Byf6sLs/003.jpg)",
      "votes": null
    },
    {
      "id": "1760668",
      "postDate": "04/19/2022 13:00:48",
      "content": "<p>Thank you for tips, I've tried save body proportions like you said and result was worst then just square resize. </p>",
      "rawMarkdown": "Thank you for tips, I've tried save body proportions like you said and result was worst then just square resize.",
      "votes": null
    },
    {
      "id": "1760670",
      "postDate": "04/19/2022 13:04:16",
      "content": "<p>I generated many 😂😂😂 DS in this competition. Letterbox is … for checking again (good idea but I think we do not use all possible pixels to provide NN information - part of image was 0) but squared bboxes works perfectly in our case. </p>",
      "rawMarkdown": "I generated many 😂😂😂 DS in this competition. Letterbox is ... for checking again (good idea but I think we do not use all possible pixels to provide NN information - part of image was 0) but squared bboxes works perfectly in our case.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1760668,
      "author_name": "kwentar",
      "author_url": "",
      "post_date": "04/19/2022 13:00:48",
      "content": "<p>Thank you for tips, I've tried save body proportions like you said and result was worst then just square resize. </p>",
      "votes": null,
      "replies": [
        {
          "id": 1760670,
          "author_name": "remekkinas",
          "author_url": "",
          "post_date": "04/19/2022 13:04:16",
          "content": "<p>I generated many 😂😂😂 DS in this competition. Letterbox is … for checking again (good idea but I think we do not use all possible pixels to provide NN information - part of image was 0) but squared bboxes works perfectly in our case. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1760661": "Congratulations to everybody! Great reaults and really fun competition. A lot of tips to learn. \nI am reading solution description and there is no new thing to add from my side. Outstanding solutions. One more tip from my side - consistent annotation to prepare data for specific input format. What does it mean?\n\nI noticed some score increase annotating dorsal fins such way they are not distorted after NN input resizing (to preserve shape of dorsal fin). All our dorsal fin were consciously annotated in form of square (w and h almost the same). The more consistent way (all dorsal fin annotated as close to square as possible) the better. As a result our dorsal fin detector was biased to square bboxes. Then when we put it to NN (before resizing to square eg. 360x360 or 768x768) we save characteristic shape of fin identity. \n\nAnnotation - as close to square as possible\n![annot](https://i.ibb.co/DWxTsFs/001.jpg)\n\nResult - no distorted dorsal fins\n![result](https://i.ibb.co/XZDjyZ2/002.jpg)\n\nAnother way was to put full body or dorsal fin using letterbox technique (commonly used in augumentation - eg. mosiac) . See below. Adding 0 pixels does not influence convolution. In this case (letterbox) more investigation is required to prove score increase.\n\n![letterbox](https://i.ibb.co/Byf6sLs/003.jpg)",
    "1760668": "Thank you for tips, I've tried save body proportions like you said and result was worst then just square resize.",
    "1760670": "I generated many 😂😂😂 DS in this competition. Letterbox is ... for checking again (good idea but I think we do not use all possible pixels to provide NN information - part of image was 0) but squared bboxes works perfectly in our case."
  },
  "source": "meta"
}