{ "id": 299457, "title": "Yolox tables and charts", "url": "/competitions/tensorflow-great-barrier-reef/discussion/299457", "author_name": "Owen Xing", "post_date": "2022-01-08T05:21:16.353000", "votes": 2, "comment_count": 0, "views": 0, "content": "

Here are the tables and charts that Yolox gives in the paper. I think it gives us a clear guide of our configuration when we adjust our model strategy.

\n
    \n

  1. Decoupled head helps the model converges quicker
    \n![https://i.stack.imgur.com/5mWES.png](url to embed)
    \nOne advantage of Yolox exceeding Yolov5 is that Yolox split classification and localization into 2 different problems, instead of together. Experiments show that doing this way helps the model converges quicker and obtain higher ap

    2. \n

  2. Data augmentation performance for Yolox-Nano and Yolox-l
    \n![https://i.stack.imgur.com/w70g0.png](url to embed)
    \nYolox absorbs great data augmentation techniques in recent years, like MixUp.

    3. \n

  3. Yolox network architecture
    \n![https://i.stack.imgur.com/1LBZN.png](url to embed)
    \nAs discussed in 1, Yolox uses decoupled head to separate cls and reg, in contrast to coupled head in Yolov3~v5, which is the reason for calling anchor-free version

    4. \n

  4. Speed and accuracy trade-off
    \n![https://i.stack.imgur.com/4HBmd.png](url to embed)
    \nThe authors of Yolox give a clear image that shows the speed and accuracy trade-off among Yolox and other Yolo series models.

    5. \n

  5. Big models and tiny models configuration
    \n![https://i.stack.imgur.com/FTxHS.png](url to embed)
    \n![https://i.stack.imgur.com/25jiV.png](url to embed)

    6. \n

  6. Backbone
    \n![https://i.stack.imgur.com/WKq5V.png](url to embed)

    7. \n
\n

Mode details are in the original paper:
\nhttps://arxiv.org/abs/2107.08430

", "messages": [ { "id": 1642145, "postDate": "2022-01-08T05:21:16.353Z", "content": "

Here are the tables and charts that Yolox gives in the paper. I think it gives us a clear guide of our configuration when we adjust our model strategy.

\n
    \n

  1. Decoupled head helps the model converges quicker
    \n![https://i.stack.imgur.com/5mWES.png](url to embed)
    \nOne advantage of Yolox exceeding Yolov5 is that Yolox split classification and localization into 2 different problems, instead of together. Experiments show that doing this way helps the model converges quicker and obtain higher ap

    2. \n

  2. Data augmentation performance for Yolox-Nano and Yolox-l
    \n![https://i.stack.imgur.com/w70g0.png](url to embed)
    \nYolox absorbs great data augmentation techniques in recent years, like MixUp.

    3. \n

  3. Yolox network architecture
    \n![https://i.stack.imgur.com/1LBZN.png](url to embed)
    \nAs discussed in 1, Yolox uses decoupled head to separate cls and reg, in contrast to coupled head in Yolov3~v5, which is the reason for calling anchor-free version

    4. \n

  4. Speed and accuracy trade-off
    \n![https://i.stack.imgur.com/4HBmd.png](url to embed)
    \nThe authors of Yolox give a clear image that shows the speed and accuracy trade-off among Yolox and other Yolo series models.

    5. \n

  5. Big models and tiny models configuration
    \n![https://i.stack.imgur.com/FTxHS.png](url to embed)
    \n![https://i.stack.imgur.com/25jiV.png](url to embed)

    6. \n

  6. Backbone
    \n![https://i.stack.imgur.com/WKq5V.png](url to embed)

    7. \n
\n

Mode details are in the original paper:
\nhttps://arxiv.org/abs/2107.08430

", "rawMarkdown": "Here are the tables and charts that Yolox gives in the paper. I think it gives us a clear guide of our configuration when we adjust our model strategy.\n\n1. Decoupled head helps the model converges quicker\n![https://i.stack.imgur.com/5mWES.png](url to embed)\nOne advantage of Yolox exceeding Yolov5 is that Yolox split classification and localization into 2 different problems, instead of together. Experiments show that doing this way helps the model converges quicker and obtain higher ap\n\n2. Data augmentation performance for Yolox-Nano and Yolox-l\n![https://i.stack.imgur.com/w70g0.png](url to embed)\nYolox absorbs great data augmentation techniques in recent years, like MixUp.\n\n3. Yolox network architecture\n![https://i.stack.imgur.com/1LBZN.png](url to embed)\nAs discussed in 1, Yolox uses decoupled head to separate cls and reg, in contrast to coupled head in Yolov3~v5, which is the reason for calling anchor-free version\n\n4. Speed and accuracy trade-off \n![https://i.stack.imgur.com/4HBmd.png](url to embed)\nThe authors of Yolox give a clear image that shows the speed and accuracy trade-off among Yolox and other Yolo series models.\n\n5. Big models and tiny models configuration\n![https://i.stack.imgur.com/FTxHS.png](url to embed)\n![https://i.stack.imgur.com/25jiV.png](url to embed)\n\n6. Backbone\n![https://i.stack.imgur.com/WKq5V.png](url to embed)\n\nMode details are in the original paper:\n[https://arxiv.org/abs/2107.08430](url)", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "1642145": "Here are the tables and charts that Yolox gives in the paper. I think it gives us a clear guide of our configuration when we adjust our model strategy.\n\n1. Decoupled head helps the model converges quicker\n![https://i.stack.imgur.com/5mWES.png](url to embed)\nOne advantage of Yolox exceeding Yolov5 is that Yolox split classification and localization into 2 different problems, instead of together. Experiments show that doing this way helps the model converges quicker and obtain higher ap\n\n2. Data augmentation performance for Yolox-Nano and Yolox-l\n![https://i.stack.imgur.com/w70g0.png](url to embed)\nYolox absorbs great data augmentation techniques in recent years, like MixUp.\n\n3. Yolox network architecture\n![https://i.stack.imgur.com/1LBZN.png](url to embed)\nAs discussed in 1, Yolox uses decoupled head to separate cls and reg, in contrast to coupled head in Yolov3~v5, which is the reason for calling anchor-free version\n\n4. Speed and accuracy trade-off \n![https://i.stack.imgur.com/4HBmd.png](url to embed)\nThe authors of Yolox give a clear image that shows the speed and accuracy trade-off among Yolox and other Yolo series models.\n\n5. Big models and tiny models configuration\n![https://i.stack.imgur.com/FTxHS.png](url to embed)\n![https://i.stack.imgur.com/25jiV.png](url to embed)\n\n6. Backbone\n![https://i.stack.imgur.com/WKq5V.png](url to embed)\n\nMode details are in the original paper:\n[https://arxiv.org/abs/2107.08430](url)" } }