{
  "id": 304414,
  "title": "Vision Transformers \"promising for future uses in object detection\" NeurIPS21GoogleBrainTeam",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/304414",
  "author_name": "Faisal Alsrheed",
  "post_date": "2022-02-01T09:54:46.976000",
  "votes": 4,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi all </p>\n<p>I just want to share this interesting paper from Google Brain Team (NeurIPS21)</p>\n<p><img src=\"https://i.postimg.cc/gkdfMvck/ViT.jpg\" alt=\"\"></p>\n<p><code>Drawing on representational similarity techniques, we find surprisingly clear differences in the features and internal structures of ViTs and CNNs.</code></p>\n<p><code>Motivated by potential future uses in object detection, we examine how well input spatial information is preserved, finding connections between spatial localization and methods of classification.</code></p>\n<p><code>We examine the potential for ViTs to be used beyond classification through a study of spatial localization, discovering ViTs with CLS tokens show strong preservation of spatial information — promising for future uses in object detection.</code></p>\n<p><a href=\"https://arxiv.org/abs/2108.08810\" target=\"_blank\">https://arxiv.org/abs/2108.08810</a></p>",
  "messages": [
    {
      "id": 1671252,
      "postDate": "2022-02-01T09:54:46.977Z",
      "content": "<p>Hi all </p>\n<p>I just want to share this interesting paper from Google Brain Team (NeurIPS21)</p>\n<p><img src=\"https://i.postimg.cc/gkdfMvck/ViT.jpg\" alt=\"\"></p>\n<p><code>Drawing on representational similarity techniques, we find surprisingly clear differences in the features and internal structures of ViTs and CNNs.</code></p>\n<p><code>Motivated by potential future uses in object detection, we examine how well input spatial information is preserved, finding connections between spatial localization and methods of classification.</code></p>\n<p><code>We examine the potential for ViTs to be used beyond classification through a study of spatial localization, discovering ViTs with CLS tokens show strong preservation of spatial information — promising for future uses in object detection.</code></p>\n<p><a href=\"https://arxiv.org/abs/2108.08810\" target=\"_blank\">https://arxiv.org/abs/2108.08810</a></p>",
      "rawMarkdown": "Hi all \n\nI just want to share this interesting paper from Google Brain Team (NeurIPS21)\n\n![](https://i.postimg.cc/gkdfMvck/ViT.jpg)\n\n\n`Drawing on representational similarity techniques, we find surprisingly clear differences in the features and internal structures of ViTs and CNNs.`\n\n`Motivated by potential future uses in object detection, we examine how well input spatial information is preserved, finding connections between spatial localization and methods of classification.`\n\n`We examine the potential for ViTs to be used beyond classification through a study of spatial localization, discovering ViTs with CLS tokens show strong preservation of spatial information — promising for future uses in object detection.`\n\n\nhttps://arxiv.org/abs/2108.08810\n",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1671252": "Hi all \n\nI just want to share this interesting paper from Google Brain Team (NeurIPS21)\n\n![](https://i.postimg.cc/gkdfMvck/ViT.jpg)\n\n\n`Drawing on representational similarity techniques, we find surprisingly clear differences in the features and internal structures of ViTs and CNNs.`\n\n`Motivated by potential future uses in object detection, we examine how well input spatial information is preserved, finding connections between spatial localization and methods of classification.`\n\n`We examine the potential for ViTs to be used beyond classification through a study of spatial localization, discovering ViTs with CLS tokens show strong preservation of spatial information — promising for future uses in object detection.`\n\n\nhttps://arxiv.org/abs/2108.08810\n"
  }
}