{
  "id": 91791,
  "title": "Training with segmentation models",
  "url": "/competitions/imaterialist-fashion-2019-FGVC6/discussion/91791",
  "author_name": "",
  "post_date": "2019-05-09T06:32:30.426101500Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hello I am new here!\nI am trying to use an usual instance segmentation model (not mask-rcnn). While designing data augmentation, I am feeling some hardness. </p>\n\n<p>It is about the 'cropsize'.\n As you well know, all images have different shape(aspect ratio) and size. I want to keep the aspect ratio of each image, but it prevents having batch size larger than 1. In order to utilize batch normalization more and faster training, it is necessary to increase the cropsize. How are you dealing with it, if you are working with segmentation models?</p>",
  "messages": [
    {
      "id": "529074",
      "postDate": "05/09/2019 06:32:30",
      "content": "<p>Hello I am new here!\nI am trying to use an usual instance segmentation model (not mask-rcnn). While designing data augmentation, I am feeling some hardness. </p>\n\n<p>It is about the 'cropsize'.\n As you well know, all images have different shape(aspect ratio) and size. I want to keep the aspect ratio of each image, but it prevents having batch size larger than 1. In order to utilize batch normalization more and faster training, it is necessary to increase the cropsize. How are you dealing with it, if you are working with segmentation models?</p>",
      "rawMarkdown": "Hello I am new here!\nI am trying to use an usual instance segmentation model (not mask-rcnn). While designing data augmentation, I am feeling some hardness. \n\nIt is about the 'cropsize'.\n As you well know, all images have different shape(aspect ratio) and size. I want to keep the aspect ratio of each image, but it prevents having batch size larger than 1. In order to utilize batch normalization more and faster training, it is necessary to increase the cropsize. How are you dealing with it, if you are working with segmentation models?",
      "votes": null
    },
    {
      "id": "534863",
      "postDate": "05/22/2019 00:07:05",
      "content": "<p>Cropping alone won't distort the image. However, in practise it will change the aspect ratio if you were to resize it to fit the network input size. In order to keep the aspect ratio in cropping, you can simply work out the positions (4 corners) to crop yourself. </p>",
      "rawMarkdown": "Cropping alone won't distort the image. However, in practise it will change the aspect ratio if you were to resize it to fit the network input size. In order to keep the aspect ratio in cropping, you can simply work out the positions (4 corners) to crop yourself.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 534863,
      "author_name": "soonyau",
      "author_url": "",
      "post_date": "05/22/2019 00:07:05",
      "content": "<p>Cropping alone won't distort the image. However, in practise it will change the aspect ratio if you were to resize it to fit the network input size. In order to keep the aspect ratio in cropping, you can simply work out the positions (4 corners) to crop yourself. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "529074": "Hello I am new here!\nI am trying to use an usual instance segmentation model (not mask-rcnn). While designing data augmentation, I am feeling some hardness. \n\nIt is about the 'cropsize'.\n As you well know, all images have different shape(aspect ratio) and size. I want to keep the aspect ratio of each image, but it prevents having batch size larger than 1. In order to utilize batch normalization more and faster training, it is necessary to increase the cropsize. How are you dealing with it, if you are working with segmentation models?",
    "534863": "Cropping alone won't distort the image. However, in practise it will change the aspect ratio if you were to resize it to fit the network input size. In order to keep the aspect ratio in cropping, you can simply work out the positions (4 corners) to crop yourself."
  },
  "source": "meta"
}