{
  "id": 307735,
  "title": "8th place solution",
  "url": "/competitions/tensorflow-great-barrier-reef/writeups/tk-8th-place-solution",
  "author_name": "",
  "post_date": "2022-02-22T02:37:30.123Z",
  "votes": 36,
  "comment_count": 10,
  "views": 0,
  "content": "<p>Congrats to all the winners, and thanks to hosts for interesting competition.</p>\n<h1>Training</h1>\n<p>data: annotated images<br>\ntrain: video_id 0, 1 (final submission is all data)<br>\nval: video_id 2 (including background images)</p>\n<h1>Models</h1>\n<h2>yolox-s</h2>\n<p>training size: 2560~3584<br>\naugmentation: default parameter<br>\ninference size: 2560, 3072, 3584, 4096<br>\ntta: WBF conf 0.05<br>\nepoch: 12<br>\ncv: 0.76</p>\n<h2>Cascade-RCNN</h2>\n<p>backbone: convnext base<br>\ntraining size: 2048~2816<br>\naugmentation: RandomGamma, CLAHE, RandomBrightnessContrast, ShiftScaleRotate, Blur,<br>\nMotionBlur, GaussNoise<br>\ninference size: 2816<br>\nepoch: 4<br>\ncv: 0.78</p>\n<h1>ensemble</h1>\n<p>WBF conf 0.2<br>\ncv: 0.795<br>\npublic: 0.611<br>\nprivate: 0.726</p>",
  "messages": [
    {
      "id": "1691464",
      "postDate": "02/15/2022 12:25:29",
      "content": "<p>Congrats to all the winners, and thanks to hosts for interesting competition.</p>\n<h1>Training</h1>\n<p>data: annotated images<br>\ntrain: video_id 0, 1 (final submission is all data)<br>\nval: video_id 2 (including background images)</p>\n<h1>Models</h1>\n<h2>yolox-s</h2>\n<p>training size: 2560~3584<br>\naugmentation: default parameter<br>\ninference size: 2560, 3072, 3584, 4096<br>\ntta: WBF conf 0.05<br>\nepoch: 12<br>\ncv: 0.76</p>\n<h2>Cascade-RCNN</h2>\n<p>backbone: convnext base<br>\ntraining size: 2048~2816<br>\naugmentation: RandomGamma, CLAHE, RandomBrightnessContrast, ShiftScaleRotate, Blur,<br>\nMotionBlur, GaussNoise<br>\ninference size: 2816<br>\nepoch: 4<br>\ncv: 0.78</p>\n<h1>ensemble</h1>\n<p>WBF conf 0.2<br>\ncv: 0.795<br>\npublic: 0.611<br>\nprivate: 0.726</p>",
      "rawMarkdown": "Congrats to all the winners, and thanks to hosts for interesting competition.\n\n# Training\n\ndata: annotated images\ntrain: video_id 0, 1 (final submission is all data)\nval: video_id 2 (including background images)\n\n# Models\n\n## yolox-s\n\ntraining size: 2560~3584\naugmentation: default parameter\ninference size: 2560, 3072, 3584, 4096\ntta: WBF conf 0.05\nepoch: 12\ncv: 0.76\n\n## Cascade-RCNN\n\nbackbone: convnext base\ntraining size: 2048~2816\naugmentation: RandomGamma, CLAHE, RandomBrightnessContrast, ShiftScaleRotate, Blur,\nMotionBlur, GaussNoise\ninference size: 2816\nepoch: 4\ncv: 0.78\n\n# ensemble\n\nWBF conf 0.2\ncv: 0.795\npublic: 0.611\nprivate: 0.726",
      "votes": null
    },
    {
      "id": "1691469",
      "postDate": "02/15/2022 12:30:11",
      "content": "<p>Looks very elegant!! Congrats on solo gold. Could you elaborate more on your training size?  Is it multiscale training from 2560 to 3584?</p>",
      "rawMarkdown": "Looks very elegant!! Congrats on solo gold. Could you elaborate more on your training size?  Is it multiscale training from 2560 to 3584?",
      "votes": null
    },
    {
      "id": "1691556",
      "postDate": "02/15/2022 13:24:18",
      "content": "<p>Thanks! I used multiscale training from 2560 to 3584.</p>",
      "rawMarkdown": "Thanks! I used multiscale training from 2560 to 3584.",
      "votes": null
    },
    {
      "id": "1691928",
      "postDate": "02/15/2022 17:33:03",
      "content": "<p>congratulations <a href=\"https://www.kaggle.com/tanakar\" target=\"_blank\">@tanakar</a> 🤩🎉🌟</p>",
      "rawMarkdown": "congratulations @tanakar 🤩🎉🌟",
      "votes": null
    },
    {
      "id": "1691994",
      "postDate": "02/15/2022 18:21:42",
      "content": "<p><code>Cascade-RCNN: \nbackbone: convnext base</code></p>\n<p>This is brilliant. Are you planning on sharing your training code for this? I would love to see how this was constructed. </p>\n<p>Either way congratulations on the gold!! </p>",
      "rawMarkdown": "`Cascade-RCNN: \nbackbone: convnext base`\n\nThis is brilliant. Are you planning on sharing your training code for this? I would love to see how this was constructed. \n\nEither way congratulations on the gold!!",
      "votes": null
    },
    {
      "id": "1692166",
      "postDate": "02/15/2022 21:37:26",
      "content": "<p>Indeed nice idea!! <br>\nI think with a \"modular\" framework as MMDet you can implement such modifications more easily, see for e.g. the public Faster-RCNN with Swin backbone, or Cascade with Resnest etc</p>",
      "rawMarkdown": "Indeed nice idea!! \nI think with a \"modular\" framework as MMDet you can implement such modifications more easily, see for e.g. the public Faster-RCNN with Swin backbone, or Cascade with Resnest etc",
      "votes": null
    },
    {
      "id": "1692322",
      "postDate": "02/16/2022 01:11:56",
      "content": "<p>I used the official convnext implementation, code is here. <a href=\"https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection\" target=\"_blank\">https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection</a></p>",
      "rawMarkdown": "I used the official convnext implementation, code is here. https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection",
      "votes": null
    },
    {
      "id": "1692336",
      "postDate": "02/16/2022 01:28:02",
      "content": "<p>Congratulations on 8th place! ConvNeXt examples look like for instance segmentation. Have you done segmentation labeling somehow or used just bounding boxes?</p>",
      "rawMarkdown": "Congratulations on 8th place! ConvNeXt examples look like for instance segmentation. Have you done segmentation labeling somehow or used just bounding boxes?",
      "votes": null
    },
    {
      "id": "1692341",
      "postDate": "02/16/2022 01:32:09",
      "content": "<p>Only bounding boxes were used. </p>",
      "rawMarkdown": "Only bounding boxes were used.",
      "votes": null
    },
    {
      "id": "1692383",
      "postDate": "02/16/2022 02:25:50",
      "content": "<p>Just Cascade rcnn can reach 0.78 with only 4 epoch on such tiny number train set?? I trained default cascade rcnn with resnet50 and train set same as yours, I trained 50 epochs only got 0.5CV 😨😨</p>",
      "rawMarkdown": "Just Cascade rcnn can reach 0.78 with only 4 epoch on such tiny number train set?? I trained default cascade rcnn with resnet50 and train set same as yours, I trained 50 epochs only got 0.5CV 😨😨",
      "votes": null
    },
    {
      "id": "1899769",
      "postDate": "08/15/2022 13:43:57",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/tanakar\" target=\"_blank\">@tanakar</a>, would you mind sharing the config that you used for this? Also, which pretrained weights did you use? Thanks!</p>",
      "rawMarkdown": "Hi @tanakar, would you mind sharing the config that you used for this? Also, which pretrained weights did you use? Thanks!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1691469,
      "author_name": "bibek777",
      "author_url": "",
      "post_date": "02/15/2022 12:30:11",
      "content": "<p>Looks very elegant!! Congrats on solo gold. Could you elaborate more on your training size?  Is it multiscale training from 2560 to 3584?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1691556,
          "author_name": "tanakar",
          "author_url": "",
          "post_date": "02/15/2022 13:24:18",
          "content": "<p>Thanks! I used multiscale training from 2560 to 3584.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1691928,
      "author_name": "arunasivapragasam",
      "author_url": "",
      "post_date": "02/15/2022 17:33:03",
      "content": "<p>congratulations <a href=\"https://www.kaggle.com/tanakar\" target=\"_blank\">@tanakar</a> 🤩🎉🌟</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1691994,
      "author_name": "trushk",
      "author_url": "",
      "post_date": "02/15/2022 18:21:42",
      "content": "<p><code>Cascade-RCNN: \nbackbone: convnext base</code></p>\n<p>This is brilliant. Are you planning on sharing your training code for this? I would love to see how this was constructed. </p>\n<p>Either way congratulations on the gold!! </p>",
      "votes": null,
      "replies": [
        {
          "id": 1692166,
          "author_name": "imeintanis",
          "author_url": "",
          "post_date": "02/15/2022 21:37:26",
          "content": "<p>Indeed nice idea!! <br>\nI think with a \"modular\" framework as MMDet you can implement such modifications more easily, see for e.g. the public Faster-RCNN with Swin backbone, or Cascade with Resnest etc</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1692322,
          "author_name": "tanakar",
          "author_url": "",
          "post_date": "02/16/2022 01:11:56",
          "content": "<p>I used the official convnext implementation, code is here. <a href=\"https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection\" target=\"_blank\">https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection</a></p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1692336,
          "author_name": "nborbit",
          "author_url": "",
          "post_date": "02/16/2022 01:28:02",
          "content": "<p>Congratulations on 8th place! ConvNeXt examples look like for instance segmentation. Have you done segmentation labeling somehow or used just bounding boxes?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 1692341,
          "author_name": "tanakar",
          "author_url": "",
          "post_date": "02/16/2022 01:32:09",
          "content": "<p>Only bounding boxes were used. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1692383,
      "author_name": "lixxxxx",
      "author_url": "",
      "post_date": "02/16/2022 02:25:50",
      "content": "<p>Just Cascade rcnn can reach 0.78 with only 4 epoch on such tiny number train set?? I trained default cascade rcnn with resnet50 and train set same as yours, I trained 50 epochs only got 0.5CV 😨😨</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1899769,
      "author_name": "chrishughes1",
      "author_url": "",
      "post_date": "08/15/2022 13:43:57",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/tanakar\" target=\"_blank\">@tanakar</a>, would you mind sharing the config that you used for this? Also, which pretrained weights did you use? Thanks!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1691464": "Congrats to all the winners, and thanks to hosts for interesting competition.\n\n# Training\n\ndata: annotated images\ntrain: video_id 0, 1 (final submission is all data)\nval: video_id 2 (including background images)\n\n# Models\n\n## yolox-s\n\ntraining size: 2560~3584\naugmentation: default parameter\ninference size: 2560, 3072, 3584, 4096\ntta: WBF conf 0.05\nepoch: 12\ncv: 0.76\n\n## Cascade-RCNN\n\nbackbone: convnext base\ntraining size: 2048~2816\naugmentation: RandomGamma, CLAHE, RandomBrightnessContrast, ShiftScaleRotate, Blur,\nMotionBlur, GaussNoise\ninference size: 2816\nepoch: 4\ncv: 0.78\n\n# ensemble\n\nWBF conf 0.2\ncv: 0.795\npublic: 0.611\nprivate: 0.726",
    "1691469": "Looks very elegant!! Congrats on solo gold. Could you elaborate more on your training size?  Is it multiscale training from 2560 to 3584?",
    "1691556": "Thanks! I used multiscale training from 2560 to 3584.",
    "1691928": "congratulations @tanakar 🤩🎉🌟",
    "1691994": "`Cascade-RCNN: \nbackbone: convnext base`\n\nThis is brilliant. Are you planning on sharing your training code for this? I would love to see how this was constructed. \n\nEither way congratulations on the gold!!",
    "1692166": "Indeed nice idea!! \nI think with a \"modular\" framework as MMDet you can implement such modifications more easily, see for e.g. the public Faster-RCNN with Swin backbone, or Cascade with Resnest etc",
    "1692322": "I used the official convnext implementation, code is here. https://github.com/facebookresearch/ConvNeXt/tree/main/object_detection",
    "1692336": "Congratulations on 8th place! ConvNeXt examples look like for instance segmentation. Have you done segmentation labeling somehow or used just bounding boxes?",
    "1692341": "Only bounding boxes were used.",
    "1692383": "Just Cascade rcnn can reach 0.78 with only 4 epoch on such tiny number train set?? I trained default cascade rcnn with resnet50 and train set same as yours, I trained 50 epochs only got 0.5CV 😨😨",
    "1899769": "Hi @tanakar, would you mind sharing the config that you used for this? Also, which pretrained weights did you use? Thanks!"
  },
  "source": "meta"
}