{
  "id": 95436,
  "title": "Is resizing images helpful?",
  "url": "/competitions/open-images-2019-object-detection/discussion/95436",
  "author_name": "",
  "post_date": "2019-06-12T07:58:42.764235500Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi, I am newbie for the object detection task, and I wonder if resizing images is helpful for training. (surely it will shorten the training time, but is it ok for inference?)\nIf so, what size of images for training/inference did you find most effective?\nI am thinking resizing all images to have longest side 800 px, but If anyone got useful insight for this, please share it for me ! :)</p>",
  "messages": [
    {
      "id": "550991",
      "postDate": "06/12/2019 07:58:42",
      "content": "<p>Hi, I am newbie for the object detection task, and I wonder if resizing images is helpful for training. (surely it will shorten the training time, but is it ok for inference?)\nIf so, what size of images for training/inference did you find most effective?\nI am thinking resizing all images to have longest side 800 px, but If anyone got useful insight for this, please share it for me ! :)</p>",
      "rawMarkdown": "Hi, I am newbie for the object detection task, and I wonder if resizing images is helpful for training. (surely it will shorten the training time, but is it ok for inference?)\nIf so, what size of images for training/inference did you find most effective?\nI am thinking resizing all images to have longest side 800 px, but If anyone got useful insight for this, please share it for me ! :)",
      "votes": null
    },
    {
      "id": "551157",
      "postDate": "06/12/2019 12:03:27",
      "content": "<p>A really important part of my process is starting small and slowly increasing the complexity of the model after I know what I am doing is right. It is better to see I am making mistakes when the model takes 10 secs to train than 5 days.</p>\n\n<p>So I start with a 128x128 models on a 1/20th of the set, then 144x, then 256x and etc.   </p>\n\n<p><a href=\"https://forums.fast.ai/t/exploring-image-size-accuracy-of-transfer-learning-in-lesson-1-pets/41724/4\">There is an interesting post on Fast.ai discussing how image size impacts error.</a> As you can see, after 160x160 on Resnet34, the error is much smaller. </p>\n\n<p><img src=\"https://forums.fast.ai/uploads/default/original/3X/e/a/eab85e535d5eba77243fd11bfe9a2d2eed74c786.png\" alt=\"\"></p>",
      "rawMarkdown": "A really important part of my process is starting small and slowly increasing the complexity of the model after I know what I am doing is right. It is better to see I am making mistakes when the model takes 10 secs to train than 5 days.\n\nSo I start with a 128x128 models on a 1/20th of the set, then 144x, then 256x and etc.   \n\n[There is an interesting post on Fast.ai discussing how image size impacts error.](https://forums.fast.ai/t/exploring-image-size-accuracy-of-transfer-learning-in-lesson-1-pets/41724/4) As you can see, after 160x160 on Resnet34, the error is much smaller. \n\n![](https://forums.fast.ai/uploads/default/original/3X/e/a/eab85e535d5eba77243fd11bfe9a2d2eed74c786.png)",
      "votes": null
    },
    {
      "id": "551826",
      "postDate": "06/13/2019 06:43:23",
      "content": "<p>thanks :) it was very helpful!</p>",
      "rawMarkdown": "thanks :) it was very helpful!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 551157,
      "author_name": "mindtrinket",
      "author_url": "",
      "post_date": "06/12/2019 12:03:27",
      "content": "<p>A really important part of my process is starting small and slowly increasing the complexity of the model after I know what I am doing is right. It is better to see I am making mistakes when the model takes 10 secs to train than 5 days.</p>\n\n<p>So I start with a 128x128 models on a 1/20th of the set, then 144x, then 256x and etc.   </p>\n\n<p><a href=\"https://forums.fast.ai/t/exploring-image-size-accuracy-of-transfer-learning-in-lesson-1-pets/41724/4\">There is an interesting post on Fast.ai discussing how image size impacts error.</a> As you can see, after 160x160 on Resnet34, the error is much smaller. </p>\n\n<p><img src=\"https://forums.fast.ai/uploads/default/original/3X/e/a/eab85e535d5eba77243fd11bfe9a2d2eed74c786.png\" alt=\"\"></p>",
      "votes": null,
      "replies": [
        {
          "id": 551826,
          "author_name": "nashory",
          "author_url": "",
          "post_date": "06/13/2019 06:43:23",
          "content": "<p>thanks :) it was very helpful!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "550991": "Hi, I am newbie for the object detection task, and I wonder if resizing images is helpful for training. (surely it will shorten the training time, but is it ok for inference?)\nIf so, what size of images for training/inference did you find most effective?\nI am thinking resizing all images to have longest side 800 px, but If anyone got useful insight for this, please share it for me ! :)",
    "551157": "A really important part of my process is starting small and slowly increasing the complexity of the model after I know what I am doing is right. It is better to see I am making mistakes when the model takes 10 secs to train than 5 days.\n\nSo I start with a 128x128 models on a 1/20th of the set, then 144x, then 256x and etc.   \n\n[There is an interesting post on Fast.ai discussing how image size impacts error.](https://forums.fast.ai/t/exploring-image-size-accuracy-of-transfer-learning-in-lesson-1-pets/41724/4) As you can see, after 160x160 on Resnet34, the error is much smaller. \n\n![](https://forums.fast.ai/uploads/default/original/3X/e/a/eab85e535d5eba77243fd11bfe9a2d2eed74c786.png)",
    "551826": "thanks :) it was very helpful!"
  },
  "source": "meta"
}