{
  "id": 302198,
  "title": "Does augmentation and higher resolution really improve LB score? 🤓",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/302198",
  "author_name": "",
  "post_date": "2022-01-21T10:17:33.224670500Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>It gets more interesting after the pandemic of training with higher resolution using yolov5. 😦</p>\n<p>By comparing of higher yolo series such as yolov5, yolox, yolor … <br>\nyolov5 only contains with augmentation methods in core code and after training with higher resolution, people are like Bazinga! </p>\n<p>I've seen other models faster-rcnn, efficientdet, retina net, DETR but it cannot outperform the yolo series </p>\n<p>Even tho this competition is small object detection, DETR might be the best suit since it is meant for small scale objection but yolov5 exceeds the limitation. </p>\n<p>So my question is &gt; <strong>Is yolov5 leading the leaderboard because of it's various augmentation methods and larger training size</strong></p>\n<p>Correct me if I'm wrong 🙌</p>",
  "messages": [
    {
      "id": "1658800",
      "postDate": "01/21/2022 10:17:33",
      "content": "<p>It gets more interesting after the pandemic of training with higher resolution using yolov5. 😦</p>\n<p>By comparing of higher yolo series such as yolov5, yolox, yolor … <br>\nyolov5 only contains with augmentation methods in core code and after training with higher resolution, people are like Bazinga! </p>\n<p>I've seen other models faster-rcnn, efficientdet, retina net, DETR but it cannot outperform the yolo series </p>\n<p>Even tho this competition is small object detection, DETR might be the best suit since it is meant for small scale objection but yolov5 exceeds the limitation. </p>\n<p>So my question is &gt; <strong>Is yolov5 leading the leaderboard because of it's various augmentation methods and larger training size</strong></p>\n<p>Correct me if I'm wrong 🙌</p>",
      "rawMarkdown": "It gets more interesting after the pandemic of training with higher resolution using yolov5. 😦\n\nBy comparing of higher yolo series such as yolov5, yolox, yolor ... \nyolov5 only contains with augmentation methods in core code and after training with higher resolution, people are like Bazinga! \n\nI've seen other models faster-rcnn, efficientdet, retina net, DETR but it cannot outperform the yolo series \n\nEven tho this competition is small object detection, DETR might be the best suit since it is meant for small scale objection but yolov5 exceeds the limitation. \n\nSo my question is > **Is yolov5 leading the leaderboard because of it's various augmentation methods and larger training size**\n\nCorrect me if I'm wrong 🙌",
      "votes": null
    },
    {
      "id": "1658809",
      "postDate": "01/21/2022 10:25:28",
      "content": "<p>public yolov5 notebooks only reach 0.62x  in LB with larger image size. <br>\nobviously top rankers have other tricks, either more data or new feature engineering or new model or combination</p>",
      "rawMarkdown": "public yolov5 notebooks only reach 0.62x  in LB with larger image size. \nobviously top rankers have other tricks, either more data or new feature engineering or new model or combination",
      "votes": null
    },
    {
      "id": "1658815",
      "postDate": "01/21/2022 10:32:30",
      "content": "<p>I guess they blend/stack/ensemble different models at the end</p>",
      "rawMarkdown": "I guess they blend/stack/ensemble different models at the end",
      "votes": null
    },
    {
      "id": "1659621",
      "postDate": "01/22/2022 02:24:03",
      "content": "<p>Here is this that I brought up about yolov5 TTA <a href=\"https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271\" target=\"_blank\">https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271</a></p>\n<p>Sampling at different resolutions might be key to picking up COTs at different sizes. I haven't done too much testing though. </p>",
      "rawMarkdown": "Here is this that I brought up about yolov5 TTA https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271\n\nSampling at different resolutions might be key to picking up COTs at different sizes. I haven't done too much testing though.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1658809,
      "author_name": "dragonzhang",
      "author_url": "",
      "post_date": "01/21/2022 10:25:28",
      "content": "<p>public yolov5 notebooks only reach 0.62x  in LB with larger image size. <br>\nobviously top rankers have other tricks, either more data or new feature engineering or new model or combination</p>",
      "votes": null,
      "replies": [
        {
          "id": 1658815,
          "author_name": "nyanswanaung",
          "author_url": "",
          "post_date": "01/21/2022 10:32:30",
          "content": "<p>I guess they blend/stack/ensemble different models at the end</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1659621,
      "author_name": "outwrest",
      "author_url": "",
      "post_date": "01/22/2022 02:24:03",
      "content": "<p>Here is this that I brought up about yolov5 TTA <a href=\"https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271\" target=\"_blank\">https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271</a></p>\n<p>Sampling at different resolutions might be key to picking up COTs at different sizes. I haven't done too much testing though. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1658800": "It gets more interesting after the pandemic of training with higher resolution using yolov5. 😦\n\nBy comparing of higher yolo series such as yolov5, yolox, yolor ... \nyolov5 only contains with augmentation methods in core code and after training with higher resolution, people are like Bazinga! \n\nI've seen other models faster-rcnn, efficientdet, retina net, DETR but it cannot outperform the yolo series \n\nEven tho this competition is small object detection, DETR might be the best suit since it is meant for small scale objection but yolov5 exceeds the limitation. \n\nSo my question is > **Is yolov5 leading the leaderboard because of it's various augmentation methods and larger training size**\n\nCorrect me if I'm wrong 🙌",
    "1658809": "public yolov5 notebooks only reach 0.62x  in LB with larger image size. \nobviously top rankers have other tricks, either more data or new feature engineering or new model or combination",
    "1658815": "I guess they blend/stack/ensemble different models at the end",
    "1659621": "Here is this that I brought up about yolov5 TTA https://www.kaggle.com/c/tensorflow-great-barrier-reef/discussion/300638#1657271\n\nSampling at different resolutions might be key to picking up COTs at different sizes. I haven't done too much testing though."
  },
  "source": "meta"
}