{"id":16949,"date":"2022-04-05T01:50:16","date_gmt":"2022-04-05T01:50:16","guid":{"rendered":"https:\/\/club.informatix.co.jp\/?p=16949"},"modified":"2024-05-03T07:16:25","modified_gmt":"2024-05-03T07:16:25","slug":"transformer%e3%82%92%e7%94%a8%e3%81%84%e3%81%9f%e9%ab%98%e7%b2%be%e5%ba%a6%e3%81%a7%e5%8a%b9%e7%8e%87%e7%9a%84%e3%81%aa%e7%89%a9%e4%bd%93%e9%a0%98%e5%9f%9f%e3%81%ae%e5%88%86%e5%89%b2%e6%89%8b%e6%b3%95","status":"publish","type":"post","link":"https:\/\/club.informatix.co.jp\/?p=16949","title":{"rendered":"Transformer\u3092\u7528\u3044\u305f\u9ad8\u7cbe\u5ea6\u3067\u52b9\u7387\u7684\u306a\u7269\u4f53\u9818\u57df\u5206\u5272\u624b\u6cd5\uff5cSegFormer"},"content":{"rendered":"<p>\u3053\u3093\u306b\u3061\u306f\u3001<a href=\"https:\/\/www.informatix.co.jp\/\" target=\"_blank\" rel=\"noopener noreferrer\">\u30a4\u30f3\u30d5\u30a9\u30de\u30c6\u30a3\u30af\u30b9<\/a>\u3067AI\u30fb\u6a5f\u68b0\u5b66\u7fd2\u696d\u52d9\u3092\u62c5\u5f53\u3057\u3066\u3044\u308b\u5927\u6a4b\u3067\u3059\u3002<\/p>\n<p>\u4eca\u56de\u306f\u753b\u50cf\u9818\u57df\u5206\u5272\uff08\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\uff09\u6280\u8853\u306e\u7d39\u4ecb\u3092\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u3053\u3053\u3067\u7d39\u4ecb\u3059\u308b\u8ad6\u6587 [1]\u306f\u3001\u8fd1\u5e74\u767a\u5c55\u8457\u3057\u3044Transformer\u3092\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306b\u53d6\u308a\u5165\u308c\u305f\u3082\u306e\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u203b\u672c\u8a18\u4e8b\u306fTransformer [2]\u3001ViT [6]\u7b49\u306e\u6df1\u5c64\u5b66\u7fd2\u306e\u77e5\u8b58\u304c\u3042\u308b\u7a0b\u5ea6\u3042\u308b\u3053\u3068\u3092\u524d\u63d0\u3068\u3057\u3066\u66f8\u3044\u3066\u3044\u307e\u3059\u3002<\/p>\n<h2>\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u3068\u306f<\/h2>\n<p>\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u753b\u50cf\u51e6\u7406\u6280\u8853\u306b\u5206\u985e\u3055\u308c\u3001\u753b\u50cf\u306e\u5404\u30d4\u30af\u30bb\u30eb\u306b\u30af\u30e9\u30b9\u3092\u5272\u308a\u5f53\u3066\u308b\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3067\u3059\u3002<\/p>\n<p>\u5177\u4f53\u7684\u306b\u306f\u753b\u50cf\u304c\u4e0e\u3048\u3089\u308c\u305f\u3068\u304d\u306b\u3001\u3069\u3053\u306b\u4eba\u304c\u5199\u3063\u3066\u3044\u3066\u3069\u3053\u306b\u8eca\u304c\u5199\u3063\u3066\u3044\u308b\u306e\u304b\u7b49\u3092\u30d4\u30af\u30bb\u30eb\u5358\u4f4d\u3067\u63a8\u8ad6\u3057\u307e\u3059\uff08\u56f31\u53f3, [3]\u3088\u308a\u629c\u7c8b\uff09\u3002<\/p>\n<p>\u753b\u50cf\u51e6\u7406\u306e\u6280\u8853\u3068\u3057\u3066\u753b\u50cf\u5206\u985e\u3084\u7269\u4f53\u691c\u77e5\u3068\u5bfe\u6bd4\u3055\u308c\u307e\u3059\u304c\u3001\u6b21\u306e\u3088\u3046\u306a\u9055\u3044\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<ul>\n<li>\u753b\u50cf\u5206\u985e\uff1a1\u679a\u306e\u753b\u50cf\u306b\u5bfe\u3057\u3066\u30af\u30e9\u30b9\u3092\u5272\u308a\u5f53\u3066\u308b\u554f\u984c<\/li>\n<li>\u7269\u4f53\u691c\u77e5\uff1a\u753b\u50cf\u306e\u4e2d\u304b\u3089\u8907\u6570\u500b\u306e\u7269\u4f53\u9818\u57df\u3092\u30dc\u30c3\u30af\u30b9\u3067\u4e88\u6e2c\u3059\u308b\u554f\u984c\uff08\u56f31\u5de6, [4]\u3088\u308a\u629c\u7c8b \uff09<\/li>\n<\/ul>\n<p>\u3064\u307e\u308a\u3001\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u3001\u3088\u308a\u8a73\u7d30\u306b\u7269\u4f53\u9818\u57df\u3092\u7279\u5b9a\u3059\u308b\u554f\u984c\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n<table style=\"border-collapse: collapse; width: 100%;\">\n<tbody>\n<tr>\n<td style=\"width: 37.6861%;\"><img decoding=\"async\" class=\"wp-image-17043 alignleft\" style=\"font-size: 16px;\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/bus_01-300x225.jpg\" alt=\"\" width=\"261\" height=\"196\" srcset=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/bus_01-300x225.jpg 300w, https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/bus_01.jpg 501w\" sizes=\"(max-width: 261px) 100vw, 261px\" \/><\/td>\n<td style=\"width: 62.3139%;\"><img decoding=\"async\" class=\"alignnone wp-image-17042\" style=\"font-size: 16px;\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/muenster00-300x149.png\" alt=\"\" width=\"393\" height=\"195\" srcset=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/muenster00-300x149.png 300w, https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/muenster00.png 510w\" sizes=\"(max-width: 393px) 100vw, 393px\" \/><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p style=\"text-align: center;\">\u56f31\uff1a\u7269\u4f53\u691c\u77e5\uff08\u5de6\uff09\u30fb\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\uff08\u53f3\uff09<\/p>\n<h2>\u5f93\u6765\u306e\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u3068\u305d\u306e\u554f\u984c\u70b9<\/h2>\n<p>\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u753b\u50cf\u5206\u985e\u6280\u8853\u3068\u95a2\u9023\u306e\u5f37\u3044\u6280\u8853\u3067\u3059\u3002<\/p>\n<p>\u753b\u50cf\u5206\u985e\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u4e00\u90e8\u3092Encoder\u3068\u3057\u3066\u5229\u7528\u3057Encoder-Decoder\u69cb\u9020\u3092\u6301\u3063\u305f\u9818\u57df\u5206\u5272\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u7d44\u307f\u7acb\u3066\u308b\u3068\u3044\u3063\u305f\u3053\u3068\u304c\u3088\u304f\u884c\u308f\u308c\u307e\u3059\u3002<\/p>\n<p>\u305d\u306e\u305f\u3081\u3001\u3053\u308c\u307e\u3067\u306f\u7573\u8fbc\u5c64\u3092\u30d9\u30fc\u30b9\u3068\u3057\u305f\u30e2\u30c7\u30eb\u304c\u4e3b\u6d41\u3067\u3057\u305f\u3002\u4f8b\u3048\u3070DeepLab [5]\u7b49\u304c\u305d\u308c\u306b\u5f53\u305f\u308a\u307e\u3059\u3002<\/p>\n<h3>\u7573\u8fbc\u5c64\u306e\u554f\u984c\u70b9<\/h3>\n<p>\u7573\u8fbc\u5c64\u306f\u5c40\u6240\u7684\u306a\u60c5\u5831\u306e\u96c6\u7d04\u3092\u7e70\u308a\u8fd4\u3059\u3053\u3068\u3067\u5f90\u3005\u306b\u305d\u306e\u53d7\u5bb9\u91ce\u3092\u5e83\u3052\u753b\u50cf\u306e\u7279\u5fb4\u91cf\u3092\u6349\u3048\u308b\u3082\u306e\u3067\u3059\u304c\u3001\u3053\u306e\u7279\u6027\u304b\u3089\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u6349\u3048\u308b\u306e\u304c\u82e6\u624b\u3068\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u96e2\u308c\u305f\u30d4\u30af\u30bb\u30eb\u306e\u60c5\u5831\u3092\u5f97\u3088\u3046\u3068\u3057\u305f\u5834\u5408\u4f55\u5c64\u3082\u7573\u8fbc\u5c64\u3092\u91cd\u306d\u306a\u3044\u3068\u3044\u3051\u307e\u305b\u3093\u30021\u3064\u306e\u7573\u8fbc\u5c64\u3067\u96c6\u7d04\u3067\u304d\u308b\u9818\u57df\u3092\u5e83\u304f\u3059\u308b\u3053\u3068\u3067\u3001\u3042\u308b\u7a0b\u5ea6\u3053\u306e\u554f\u984c\u306f\u7de9\u548c\u3067\u304d\u307e\u3059\u3002<\/p>\n<p>\u4f8b\u3048\u3070DeepLab\u7b49\u3067\u7528\u3044\u3089\u308c\u3066\u3044\u308bAtrous Convolution\u304c\u3053\u306e\u65b9\u9762\u306e\u6539\u5584\u306b\u3042\u305f\u308a\u307e\u3059\u3002\u3057\u304b\u3057\u3001\u6839\u672c\u7684\u306b\u89e3\u6c7a\u3059\u308b\u8a33\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002<\/p>\n<p>\u753b\u50cf\u5206\u985e\u306f\u6bb5\u968e\u7684\u306b\u753b\u50cf\u306e\u30b5\u30a4\u30ba\u3092\u5c0f\u3055\u304f\u3057\u3066\u3044\u304d\u6700\u7d42\u7684\u306b\u753b\u50cf1\u679a\u306b\u5bfe\u3057\u3066\u30af\u30e9\u30b9\u3092\u4e88\u6e2c\u3057\u307e\u3059\u3002\u305d\u306e\u305f\u3081\u6700\u7d42\u7684\u306b\u306f\u5c40\u6240\u7684\u306a\u60c5\u5831\u3068\u5927\u57df\u7684\u306a\u60c5\u5831\u306f\u878d\u5408\u3055\u308c\u307e\u3059\u3002<\/p>\n<p>\u4e00\u65b9\u3067\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u30d4\u30af\u30bb\u30eb\u5358\u4f4d\u3067\u306e\u4e88\u6e2c\u306b\u306a\u308b\u305f\u3081\u3001\u30e2\u30c7\u30eb\u306e\u51fa\u529b\u306f\u5165\u529b\u753b\u50cf\u3068\u540c\u7a0b\u5ea6\u306e\u89e3\u50cf\u5ea6\u306b\u306a\u308a\u307e\u3059\u3002\u3064\u307e\u308a\u3001\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306e\u753b\u50cf\u5206\u985e\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u5229\u7528\u3059\u308b\u5834\u5408\u3001\u3069\u3053\u304b\u3067\u62e1\u5927\u64cd\u4f5c\u3092\u3057\u306a\u3044\u3068\u3044\u3051\u306a\u304f\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\u305d\u306e\u5834\u5408\u3001\u5c40\u6240\u7684\u306a\u64cd\u4f5c\u3092\u884c\u3046\u7573\u8fbc\u5c64\u306e\u307f\u3067\u306f\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u6349\u3048\u3065\u3089\u304f\u306a\u308a\u3001\u826f\u3044\u7cbe\u5ea6\u3092\u51fa\u3059\u305f\u3081\u306b\u306f\u8907\u96d1\u306aDecoder\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306f\u96e2\u308c\u305f\u30d4\u30af\u30bb\u30eb\u3069\u3046\u3057\u304c\u540c\u3058\u7269\u4f53\u306a\u306e\u304b\u3092\u753b\u50cf\u306e\u5927\u57df\u7684\u60c5\u5831\u304b\u3089\u5224\u65ad\u3057\u306a\u3044\u3068\u3044\u3051\u306a\u3044\u3053\u3068\u3092\u9451\u307f\u308b\u306b\u3001\u7573\u8fbc\u5c64\u3092\u57fa\u672c\u3068\u3057\u305f\u69cb\u9020\u306f\u6700\u9069\u3067\u306f\u306a\u3044\u3088\u3046\u306b\u601d\u3048\u307e\u3059\u3002<\/p>\n<h3>Transformer\u3092\u7528\u3044\u305f\u5148\u884c\u7814\u7a76\u306e\u554f\u984c\u70b9<\/h3>\n<p>\u3053\u306e\u3088\u3046\u306a\u72b6\u6cc1\u306fViT [6]\u306e\u767b\u5834\u3067\u5927\u304d\u304f\u5909\u308f\u308a\u307e\u3057\u305f\u3002<\/p>\n<p>ViT\u306f<a href=\"https:\/\/club.informatix.co.jp\/?p=922\" target=\"_blank\" rel=\"noopener\">\u81ea\u7136\u8a00\u8a9e\u51e6\u7406<\/a>\u3067\u7528\u3044\u3089\u308c\u3066\u3044\u308bTransformer\u3068\u547c\u3070\u308c\u308b\u6280\u8853\u3092\u753b\u50cf\u5206\u985e\u306b\u5fdc\u7528\u3057\u305f\u3082\u306e\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n<p>\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u3068\u5927\u304d\u304f\u7570\u306a\u308b\u70b9\u306f\u3001\uff08\u3082\u3061\u308d\u3093\u8a2d\u8a08\u306b\u3088\u308a\u307e\u3059\u304c\uff09Transformer\u5c64\u306f1\u5c64\u3067\u753b\u50cf\u306e\u5168\u9818\u57df\u304b\u3089\u60c5\u5831\u3092\u76f4\u63a5\u53d6\u308a\u8fbc\u3081\u308b\u3053\u3068\u3067\u3059\u3002<\/p>\n<p>ViT\u306f\u753b\u50cf\u5206\u985e\u306e\u30e2\u30c7\u30eb\u3067\u3059\u304c\u3001\u3053\u308c\u3092\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306eEncoder\u306b\u8ee2\u7528\u3057\u305f\u30e2\u30c7\u30eb\u304c\u591a\u6570\u7814\u7a76\u30fb\u767a\u8868\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u4ee3\u8868\u7684\u306a\u3082\u306e\u3068\u3057\u3066\u306fSETR [7]\u304c\u6319\u3052\u3089\u308c\u307e\u3059\u3002<\/p>\n<p>Transformer\u3092\u30d9\u30fc\u30b9\u3068\u3057\u305f\u30e2\u30c7\u30eb\u306f\u5148\u306b\u8ff0\u3079\u305f\u53d7\u5bb9\u91ce\u306e\u554f\u984c\u3092\u6839\u672c\u304b\u3089\u89e3\u6c7a\u3059\u308b\u3082\u306e\u3067\u3059\u304c\u3001\u554f\u984c\u70b9\u304c\u3044\u304f\u3064\u304b\u3042\u308a\u307e\u3059\u3002<\/p>\n<p>\u4e00\u3064\u76ee\u306fTransformer\u7279\u6709\u306e\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u306b\u95a2\u3059\u308b\u3082\u306e\u3067\u3059\u3002\u901a\u5e38\u306e\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u306f\u56fa\u5b9a\u9577\u306e\u305f\u3081\u3001\u5b66\u7fd2\u6642\u3068\u4e88\u6e2c\u6642\u3067\u753b\u50cf\u30b5\u30a4\u30ba\u304c\u7570\u306a\u308b\u5834\u5408\u3046\u307e\u304f\u5bfe\u5fdc\u3067\u304d\u307e\u305b\u3093\u3002<\/p>\n<p>\u4e8c\u3064\u76ee\u306fAttention\u6a5f\u69cb\u306e\u30e1\u30e2\u30ea\u30fc\u6d88\u8cbb\u3084\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u5927\u304d\u3044\u3053\u3068\u3067\u3059\u3002<\/p>\n<p>SETR\u306e\u5834\u5408\u306fViT\u3092\u305d\u306e\u307e\u307eEncoder\u3068\u3057\u3066\u7528\u3044\u3066\u3044\u308b\u305f\u3081\u3001\u5404Transformer\u5c64\u306e\u51fa\u529b\u304c\u540c\u3058\u7a7a\u9593\u89e3\u50cf\u5ea6\u3092\u6301\u3063\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306f\u968e\u5c64\u69cb\u9020\u3092\u6301\u3063\u3066\u3044\u308b\u30e2\u30c7\u30eb\u306b\u6bd4\u3079\u3066\u8a08\u7b97\u91cf\u3092\u5897\u5927\u3055\u305b\u308b\u539f\u56e0\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u7279\u5fb4\u91cf\u306e\u7a7a\u9593\u89e3\u50cf\u5ea6\u306f\u5143\u753b\u50cf\u306e\\(\\frac{1}{16}\\)\u3068\u4f4e\u89e3\u50cf\u5ea6\u3067\u3001\u30d4\u30af\u30bb\u30eb\u5358\u4f4d\u306e\u8a73\u7d30\u306a\u4e88\u6e2c\u3092\u3059\u308b\u3068\u3044\u3046\u89b3\u70b9\u304b\u3089\u3082\u826f\u3044\u69cb\u9020\u3068\u306f\u8a00\u3048\u306a\u305d\u3046\u3067\u3059\u3002\u307e\u305f\u3001SETR\u306f\u3044\u304f\u3064\u3082\u306e\u7573\u8fbc\u5c64\u3092\u5229\u7528\u3057\u8a08\u7b97\u30b3\u30b9\u30c8\u306e\u5927\u304d\u3044\u30c7\u30b3\u30fc\u30c9\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u4ee5\u4e0a\u306e\u3053\u3068\u3092\u8e0f\u307e\u3048\u3001Transformer\u306e\u30e2\u30c7\u30eb\u69cb\u9020\u3092\u63f4\u7528\u3057\u3001\u65e2\u5b58\u306eTransformer\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u306b\u6bd4\u3079\u3066\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u4e0b\u3052\u3064\u3064\u7cbe\u5ea6\u5411\u4e0a\u3092\u76ee\u6307\u3057\u305fSegFormer\u3068\u3044\u3046\u30e2\u30c7\u30eb\u3092\u63d0\u6848\u3057\u305f\u306e\u304c\u3001\u7d39\u4ecb\u3059\u308b\u8ad6\u6587\u306e\u5185\u5bb9\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n<h2>\u30e2\u30c7\u30eb\u6982\u8981<\/h2>\n<p>SegFormer\u306e\u6982\u8981\u306f\u4ee5\u4e0b\u306b\u306a\u308a\u307e\u3059\u3002\u8a08\u7b97\u3092\u52b9\u7387\u5316\u3057\u305fTransformer\u3092\u30d9\u30fc\u30b9\u306bEncoder-Decoder\u69cb\u9020\u3092\u3057\u305f\u3082\u306e\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p><img decoding=\"async\" class=\"wp-image-17051 aligncenter\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/model.png\" alt=\"\" width=\"559\" height=\"282\" srcset=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/model.png 644w, https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/model-300x151.png 300w\" sizes=\"(max-width: 559px) 100vw, 559px\" \/><\/p>\n<p style=\"text-align: center;\">\u56f32\uff1aSegFormer\uff08[1]\u3088\u308a\u629c\u7c8b\uff09<\/p>\n<p>SegFormer\u306f\u7cbe\u5ea6\u5411\u4e0a\u3068\u8a08\u7b97\u30b3\u30b9\u30c8\u8efd\u6e1b\u306e\u305f\u3081\u306b\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u3044\u304f\u3064\u304b\u306e\u5de5\u592b\u304c\u53d6\u308a\u5165\u308c\u3089\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n<ol>\n<li>\u30d1\u30c3\u30c1\u57cb\u3081\u8fbc\u307f\u306e\u5de5\u592b<\/li>\n<li>Attention\u6a5f\u69cb\u306e\u52b9\u7387\u5316<\/li>\n<li>Positional Encoding\u3092\u52d5\u7684\u306b\u8a08\u7b97<\/li>\n<li>\u7279\u5fb4\u91cf\u306e\u968e\u5c64\u69cb\u9020\u5316<\/li>\n<li>Decoder\u306e\u7c21\u7565\u5316<\/li>\n<\/ol>\n<p>\u305d\u308c\u305e\u308c\u4ee5\u4e0b\u306b\u8aac\u660e\u3057\u307e\u3059\u3002<\/p>\n<h2>\u5de5\u592b\u8a73\u7d30<\/h2>\n<h3>1. \u30d1\u30c3\u30c1\u57cb\u3081\u8fbc\u307f\u306e\u5de5\u592b<\/h3>\n<p>\u901a\u5e38\u306eViT\u306f\u5165\u529b\u753b\u50cf\u309216\u00d716\u3092\uff11\u30d1\u30c3\u30c1\u3068\u3057\u3066\u7279\u5fb4\u91cf\u3092\u69cb\u6210\u3092\u3057\u3066\u3044\u307e\u3059\u304c\u3001SegFormer\u3067\u306f4\u00d74\u3092\uff11\u30d1\u30c3\u30c1\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3053\u308c\u306f\u3001ViT\u306f\u753b\u50cf\u5206\u985e\u306e\u30e2\u30c7\u30eb\u3067\u753b\u50cf\u3092\u5927\u96d1\u628a\u306b\u6349\u3048\u308c\u3070\u826f\u3044\u306e\u306b\u5bfe\u3057\u3066\u3001SegFormer\u306f\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306e\u30e2\u30c7\u30eb\u3067\u3088\u308a\u8a73\u7d30\u306b\u753b\u50cf\u3092\u6349\u3048\u306a\u3051\u308c\u3070\u3044\u3051\u306a\u3044\u305f\u3081\u3067\u3059\u3002\u753b\u50cf\u306e\u5165\u529b\u6642\u306b\u3001\u7a7a\u9593\u65b9\u5411\u306e\u60c5\u5831\u3092\u4e0d\u5fc5\u8981\u306b\u6f70\u3055\u306a\u3044\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u307e\u305f\u3001ViT\u7b49\u3067\u306f\u57cb\u3081\u8fbc\u3080\u969b\u306b\u30d1\u30c3\u30c1\u306e\u91cd\u306a\u308a\u304c\u7121\u3044\u3088\u3046\u306b\u3057\u3066\u3044\u307e\u3059\u304cSegFormer\u3067\u306f\u3042\u308b\u7a0b\u5ea6\u306e\u91cd\u306a\u308a\u3092\u6301\u305f\u305b\u3066\u57cb\u3081\u8fbc\u3093\u3067\u3044\u307e\u3059\u3002\u3053\u308c\u306f\u3001\u8a08\u7b97\u91cf\u304c\u5897\u3048\u3066\u3057\u307e\u3044\u307e\u3059\u304c\u8fd1\u96a3\u306e\u30d1\u30c3\u30c1\u306e\u7279\u5fb4\u91cf\u306e\u9023\u7d9a\u6027\u3092\u62c5\u4fdd\u3057\u3001\u7cbe\u5ea6\u3092\u5411\u4e0a\u3055\u305b\u308b\u3053\u3068\u3092\u76ee\u7684\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u53c2\u8003\u8ad6\u6587\u3067\u306f2\u6b21\u5143\u306e\u7573\u8fbc\u5c64\u3067\u30d1\u30c3\u30c1\u57cb\u3081\u8fbc\u307f\u3092\u5b9f\u73fe\u3057\u3066\u304a\u308a\u3001\\(K,S,P\\)\u3092\u305d\u308c\u305e\u308c\u30ab\u30fc\u30cd\u30eb\u30b5\u30a4\u30ba\u3001\u30b9\u30c8\u30e9\u30a4\u30c9\u3001\u30d1\u30c7\u30a3\u30f3\u30b0\u3068\u3059\u308b\u3068\u3001\u5165\u529b\u753b\u50cf\u306e\u30d1\u30c3\u30c1\u57cb\u3081\u8fbc\u307f\u306b\u5bfe\u3057\u3066\u306f\\(K=7,S=4,P=3\\)\u3001\u5404\u30d6\u30ed\u30c3\u30af\u9593\u3067\u306e\u30d1\u30c3\u30c1\u57cb\u3081\u8fbc\u307f\u3067\u306f\\(K=3,S=2,P=1\\)\u3068\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<h3>2. Attention\u6a5f\u69cb\u306e\u52b9\u7387\u5316<\/h3>\n<p>Attention\u6a5f\u69cb\u306f\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u76f4\u63a5\u6349\u3048\u307e\u3059\u304c\u3001\u305d\u306e\u4ee3\u511f\u3068\u3057\u3066\u8a08\u7b97\u30b3\u30b9\u30c8\u304c\u5927\u304d\u3044\u3067\u3059\u3002\u305d\u306e\u305f\u3081\u3001Attention\u6a5f\u69cb\u306e\u52b9\u7387\u5316\u306f\u91cd\u8981\u306a\u554f\u984c\u3067\u3059\u3002<\/p>\n<p>\u53c2\u8003\u8ad6\u6587\u3067\u306f\u3001Attention\u6a5f\u69cb\u306e\u69cb\u9020\u306f\u305d\u306e\u307e\u307e\u306b\u9069\u5207\u306b\u6b21\u5143\u3092\u4e0b\u3052\u308b\u3053\u3068\u306b\u3088\u3063\u3066\u8a08\u7b97\u30b3\u30b9\u30c8\u306e\u524a\u6e1b\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002Attention\u6a5f\u69cb\u306f\u3001\u5165\u529b\u9577\\(N\\)\u3001\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\\(C\\)\u3068\u3057\u305f\u6642\u3001\\((N, C)\\)\u306e\u6b21\u5143\u3092\u6301\u3063\u305f\u30c6\u30f3\u30bd\u30ebKey(\\(K\\))\u3001Value(\\(V\\))\u3068Query(\\(Q\\))\u306b\u5bfe\u3057\u3066<\/p>\n<p>\\[Attention(Q,K,V)=Softmax(\\frac{QK^{T}}{\\sqrt{C}})V\\]<\/p>\n<p>\u3068\u8868\u3055\u308c\u307e\u3059\u3002\u3053\u306e\u8a08\u7b97\u306e\u8907\u96d1\u5ea6\u306f\\(\\mathcal{O}(N^2)\\)\u3067\u3059\u3002\u3053\u306e\\(K,V\\)\u306b\u5bfe\u3057\u3001\u30c6\u30f3\u30bd\u30eb\u5909\u5f62(\\(Reshape\\))\u3068\u7dda\u5f62\u5909\u63db(\\(Linear\\))\u3092\u7528\u3044\u3066\u5165\u529b\u9577\u3092\u5727\u7e2e\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3064\u307e\u308a\u3001\u5727\u7e2e\u7387\u3092\\(R\\)\u3068\u3057\u305f\u3068\u304d\u306b\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u5909\u63db\u3092\u65bd\u3057\u307e\u3059\u3002<\/p>\n<p>\\[\\hat{K}=Reshape(\\frac{N}{R},C*R)(K)\\]<\/p>\n<p>\\[K=Linear(C*R,C)(\\hat{K})\\]<\/p>\n<p>\u3053\u306e\u3088\u3046\u306b\u3059\u308b\u3053\u3068\u3067\\(K,V\\)\u306f\\((\\frac{N}{R},C)\\)\u306e\u6b21\u5143\u3092\u6301\u3061\u3001Attention\u306e\u8a08\u7b97\u306e\u8907\u96d1\u5ea6\u304c\\(\\mathcal{O}(\\frac{N^2}{R})\\)\u306b\u524a\u6e1b\u3055\u308c\u307e\u3059\u3002<\/p>\n<h3>3. Positional Encoding\u3092\u52d5\u7684\u306b\u8a08\u7b97\uff08Mix-FFN\uff09<\/h3>\n<p>Attention\u6a5f\u69cb\u3067\u306f\u4f4d\u7f6e\u60c5\u5831\u304c\u53d6\u308a\u8fbc\u307e\u308c\u306a\u3044\u305f\u3081\u3001\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f(Positional Encoding)\u3068\u547c\u3070\u308c\u308b\u30c6\u30f3\u30bd\u30eb\u306e\u52a0\u7b97\u304c\u884c\u308f\u308c\u307e\u3059\u3002<\/p>\n<p>\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u306f\u901a\u5e38\u56fa\u5b9a\u30b5\u30a4\u30ba\u306e\u305f\u3081\u3001\u5b66\u7fd2\u6642\u3068\u4e88\u6e2c\u6642\u3067\u753b\u50cf\u30b5\u30a4\u30ba\u304c\u9055\u3046\u5834\u5408\u306b\u5bfe\u5fdc\u3067\u304d\u307e\u305b\u3093\u3002<\/p>\n<p>\u4e88\u6e2c\u6642\u306e\u753b\u50cf\u30b5\u30a4\u30ba\u306b\u5408\u308f\u305b\u3066\u9069\u5207\u306b\u30ea\u30b5\u30a4\u30ba\u3092\u884c\u3046\u3053\u3068\u306b\u3088\u3063\u3066\u5bfe\u5fdc\u3059\u308b\u3053\u3068\u3082\u3067\u304d\u307e\u3059\u304c\u3001\u305d\u308c\u304c\u7cbe\u5ea6\u52a3\u5316\u3092\u62db\u304f\u3053\u3068\u3082\u3042\u308b\u3088\u3046\u3067\u3059\u3002<\/p>\n<p>\u3053\u306e\u3088\u3046\u306a\u554f\u984c\u306b\u5bfe\u5fdc\u3059\u308b\u305f\u3081\u53c2\u8003\u8ad6\u6587\u3067\u306f\u3001\u56fa\u5b9a\u30b5\u30a4\u30ba\u306e\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u3092\u6368\u3066\u3066\u52d5\u7684\u306b\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f\u3092\u8a08\u7b97\u3067\u304d\u308b\u3088\u3046\u306b\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u5177\u4f53\u7684\u306b\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u591a\u91cd\u7dda\u5f62\u5c64(\\(MLP\\))\u3001\u7573\u8fbc\u5c64(\\(Conv\\))\u3068GELU\u6d3b\u6027\u5316\u95a2\u6570(\\(GELU\\))\u306e\u7d44\u307f\u5408\u308f\u305b\u3067\u8a08\u7b97\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\\[ x_{out} = MLP(GELU(Conv_{3\\times 3}(MLP(x_{in}))))+x_{in}\\]<\/p>\n<p>\u3053\u3053\u3067\u3001\\(MLP(GELU(Conv_{3\\times 3}(MLP(x_{in}))))\\)\u90e8\u5206\u304c\u52d5\u7684\u306aPositional Encoding\u306b\u5bfe\u5fdc\u3057\u3066\u304a\u308a\u3001\\(Conv_{3\\times 3}\\)\u3067\u76f8\u5bfe\u7684\u306a\u4f4d\u7f6e\u60c5\u5831\u306e\u96c6\u7d04\u4f5c\u696d\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<h3>4. \u7279\u5fb4\u91cf\u306e\u968e\u5c64\u69cb\u9020\u5316<\/h3>\n<p>\u753b\u50cf\u5206\u985e\u30e2\u30c7\u30eb\u3067\u3042\u308bViT\u3084\u305d\u308c\u3092\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u306b\u76f4\u63a5\u5fdc\u7528\u3057\u305fSETR\u3067\u306f\u3001Transformer\u5404\u5c64\u306e\u51fa\u529b\u3059\u308b\u7279\u5fb4\u91cf\u306e\u7a7a\u9593\u89e3\u50cf\u5ea6\u306f\u5909\u308f\u308a\u307e\u305b\u3093\u3002<\/p>\n<p>\u53c2\u8003\u8ad6\u6587\u3067\u306f\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u540c\u69d8\u3001\u5404\u30d6\u30ed\u30c3\u30af\u3054\u3068\u306b\u89e3\u50cf\u5ea6\u3092\u4e0b\u3052\u3066\u3044\u304f\u69cb\u9020\u3092\u63a1\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u3053\u308c\u306f\u3001\u89e3\u50cf\u5ea6\u3092\u4e0b\u3052\u308b\u3053\u3068\u3067\u8a08\u7b97\u30b3\u30b9\u30c8\u3092\u4e0b\u3052\u308b\u3068\u3068\u3082\u306b\u3001\u9ad8\u89e3\u50cf\u5ea6\u306e\u7279\u5fb4\u91cf\u306f\u3088\u308a\u8a73\u7d30\u306a\u5c40\u6240\u7684\u60c5\u5831\u306b\u7126\u70b9\u3092\u5f53\u3066\u3001\u4f4e\u89e3\u50cf\u5ea6\u306e\u7279\u5fb4\u91cf\u306f\u3088\u308a\u5927\u57df\u7684\u60c5\u5831\u306b\u7126\u70b9\u3092\u5f53\u3066\u308b\u3053\u3068\u3092\u76ee\u7684\u306b\u3057\u3066\u3044\u307e\u3059\u3002\u5177\u4f53\u7684\u306b\u306f\u30d1\u30c3\u30c1\u306e\u5de5\u592b\u3067\u8ff0\u3079\u305f2\u6b21\u5143\u7573\u8fbc\u5c64\u3067\u884c\u3044\u307e\u3059\u3002<\/p>\n<h3>5. Decoder\u306e\u7c21\u7565\u5316<\/h3>\n<p>\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u306e\u5834\u5408Encoder\u304c\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u6349\u3048\u304d\u308c\u306a\u3044\u305f\u3081\u3001Decoder\u3082\u7573\u8fbc\u5c64\u3092\u7528\u3044\u3066\u5468\u8fba\u60c5\u5831\u3092\u5229\u7528\u3059\u308b\u65b9\u304c\u6709\u5229\u306a\u305f\u3081\u5fc5\u7136\u7684\u306b\u8907\u96d1\u5316\u3057\u307e\u3059\u3002<\/p>\n<p>\u4e00\u65b9\u3067SegFormer\u306f\u30a8\u30f3\u30b3\u30fc\u30c9\u3057\u305f\u60c5\u5831\u81ea\u4f53\u304c\u65e2\u306b\u5927\u57df\u7684\u60c5\u5831\u3092\u5341\u5206\u306b\u6349\u3048\u3066\u3044\u308b\u306f\u305a\u3068\u3044\u3046\u4e3b\u5f35\u306e\u3082\u3068\u3001\u7c21\u7d20\u306aDecoder\u3092\u63a1\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u5177\u4f53\u7684\u306b\u306f\u30014\u3064\u306e\u5404Transformer\u30d6\u30ed\u30c3\u30af\\(i\\in \\{ 1,2,3,4\\} \\)\u304b\u3089\u51fa\u529b\u3055\u308c\u308b\u7279\u5fb4\u91cf\\(F_{i}\\)\u3092\u3001\u9069\u5207\u306b\u5f62\u3092\u6574\u3048\u3066\u7d50\u5408\u30fb\u7dda\u5f62\u5909\u63db\u3059\u308b\u3060\u3051\u306e\u3082\u306e\u3067\u3059\u3002<\/p>\n<p>\u5404\\(F_{i}\\)\u306f\u7a7a\u9593\u89e3\u50cf\u5ea6\\(\\frac{H}{2^{i+1}}\\times \\frac{W}{2^{i+1}}\\)\u3001\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\\(C_{i}\\)\u3092\u6301\u3061\u307e\u3059\u3002\u305f\u3060\u3057\\(C_{i+1}\\geq C_{i}\\)\u3002\u3053\u308c\u3089\u3092\u5168\u3066\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u7dda\u5f62\u5909\u63db\u3067\u5171\u901a\u306e\u30c1\u30e3\u30f3\u30cd\u30eb\u6570\\(C\\)\u306b\u7d71\u4e00\u3057\u3001\u307e\u305f\u5171\u901a\u306e\u7a7a\u9593\u89e3\u50cf\u5ea6\\(\\frac{H}{4}\\times \\frac{W}{4}\\)\u306b\u62e1\u5927\u3057\u307e\u3059\u3002<\/p>\n<p>\u305d\u3057\u3066\u30c1\u30e3\u30f3\u30cd\u30eb\u65b9\u5411\u306b\u7d50\u5408\u3057\u3066\u3001\u3055\u3089\u306b\u7dda\u5f62\u5909\u63db\u3092\u3057\u307e\u3059\u3002\u3053\u306e\u69d8\u306a\u7c21\u7d20\u306a\u69cb\u9020\u306f\u8a08\u7b97\u91cf\u524a\u6e1b\u3092\u610f\u56f3\u3057\u3066\u3044\u307e\u3059\u3002\\(N_{cls}\\)\u306f\u5bfe\u8c61\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u30af\u30e9\u30b9\u6570\u3067\u3059\u3002<\/p>\n<p>\\[\\hat{F_{i}}=Linear(C_{i},C)(F_{i}) ,\\forall i\\]<\/p>\n<p>\\[\\hat{F_{i}}=Upsample(\\frac{H}{4}\\times \\frac{W}{4})(\\hat{F_{i}}) ,\\forall i\\]<\/p>\n<p>\\[F=Linear(4C,C)(Concat(\\hat{F_{i}})) ,\\forall i\\]<\/p>\n<p>\\[M=Linear(C,N_{cls})(F)\\]<\/p>\n<h2>\u7cbe\u5ea6<\/h2>\n<h3>\u5b66\u7fd2\u6982\u8981<\/h3>\n<p>SegFormer\u306e\u5b66\u7fd2\u30fb\u7cbe\u5ea6\u8a55\u4fa1\u306fCityscapes [3]\u3001ADE20K [8]\u3001COCO stuff [9]\u3068\u3044\u30463\u3064\u306e\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u3067\u884c\u308f\u308c\u3066\u3044\u307e\u3059\u3002Encoder\u306fImagenet-1K\u3067\u4e8b\u524d\u5b66\u7fd2\u3092\u884c\u3044\u3001Decoder\u306f\u30e9\u30f3\u30c0\u30e0\u306b\u521d\u671f\u5316\u3092\u3057\u305f\u3046\u3048\u3067\u5b66\u7fd2\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u30c7\u30fc\u30bf\u62e1\u5f35\u3068\u3057\u3066\u306f\u500d\u7387\\(0.5\\sim 2\\)\u306e\u30e9\u30f3\u30c0\u30e0\u30ea\u30b5\u30a4\u30ba\u3001\u6c34\u5e73\u65b9\u5411\u306e\u30e9\u30f3\u30c0\u30e0\u53cd\u8ee2\u3001\u30e9\u30f3\u30c0\u30e0\u306e\u5207\u308a\u629c\u304d\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>SegFormer\u306f\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u5c0f\u3055\u3044\u9806\u306bMiT-B0\u304b\u3089MiT-B5\u307e\u3067\u8a66\u3057\u3066\u3044\u307e\u3059\u3002\u30e2\u30c7\u30eb\u306e\u8a73\u7d30\u306f\u8ad6\u6587 [1]\u3092\u53c2\u7167\u3057\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<h3>\u7cbe\u5ea6<\/h3>\n<p>ADE20K\u306b\u5bfe\u3057\u3066\u4ee3\u8868\u7684\u306a\u30e2\u30c7\u30eb\u3068\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u6570\u30fb\u7cbe\u5ea6\u3092\u8868\u3057\u305f\u306e\u304c\u4ee5\u4e0b\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n<p><img decoding=\"async\" class=\"wp-image-17105 aligncenter\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/perform.png\" alt=\"\" width=\"448\" height=\"400\" srcset=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/perform.png 564w, https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/perform-300x268.png 300w\" sizes=\"(max-width: 448px) 100vw, 448px\" \/><\/p>\n<p style=\"text-align: center;\">\u56f33\uff1a\u8a08\u7b97\u52b9\u7387\u6bd4\u8f03\uff08[1]\u3088\u308a\u629c\u7c8b\uff09<\/p>\n<p>\u8a2d\u8a08\u306e\u76ee\u8ad6\u898b\u901a\u308a\u3001\u4ed6\u306e\u4ee3\u8868\u7684\u306a\u30e2\u30c7\u30eb\u3068\u6bd4\u3079\u3066\u5c11\u30d1\u30e9\u30e1\u30fc\u30bf\u3067\u9ad8\u7cbe\u5ea6\u304c\u51fa\u3066\u3044\u308b\u3053\u3068\u304c\u78ba\u8a8d\u3067\u304d\u307e\u3059\u3002<\/p>\n<p>ADE20K\u3068Cityscapes\u306b\u5bfe\u3059\u308b\u8a73\u7d30\u306a\u7d50\u679c\u6bd4\u8f03\u306f\u4ee5\u4e0b\u3067\u3059\u3002<\/p>\n<p><img decoding=\"async\" class=\"alignnone size-full wp-image-17058\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/precision.png\" alt=\"\" width=\"758\" height=\"419\" srcset=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/precision.png 758w, https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/precision-300x166.png 300w\" sizes=\"(max-width: 758px) 100vw, 758px\" \/><\/p>\n<p style=\"text-align: center;\">\u56f34\uff1a\u7cbe\u5ea6\u6bd4\u8f03\uff08[1]\u3088\u308a\u629c\u7c8b\uff09<\/p>\n<p>\u901f\u5ea6\u9762\u3067\u306f\u82e5\u5e72\u898b\u52a3\u308a\u3059\u308b\u90e8\u5206\u306f\u3042\u308a\u307e\u3059\u304c\u3001\u3044\u305a\u308c\u306e\u30b1\u30fc\u30b9\u306b\u304a\u3044\u3066\u3082\u5c11\u30d1\u30e9\u30e1\u30fc\u30bf\u3067\u9ad8\u7cbe\u5ea6\u304c\u51fa\u3066\u3044\u308b\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n<h3>\u5dee\u5206\u691c\u8a3c<\/h3>\n<p>\u5404\u5de5\u592b\u304c\u7cbe\u5ea6\u306b\u3069\u308c\u3060\u3051\u52b9\u3044\u3066\u3044\u308b\u304b\u3082\u691c\u8a3c\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u5148\u305a\u306f\u52d5\u7684\u306a\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f(Mix-FFN)\u304c\u56fa\u5b9a\u9577\u306e\u4f4d\u7f6e\u57cb\u3081\u8fbc\u307f(PE)\u306b\u5bfe\u3057\u3066\u512a\u4f4d\u6027\u3092\u6301\u3063\u3066\u3044\u308b\u304b\u3092Cityscapes\u3067\u78ba\u8a8d\u3057\u3066\u3044\u307e\u3059\u3002Cityscapes\u3067\u306f\\(1024\\times 1024\\)\u306e\u753b\u50cf\u30b5\u30a4\u30ba\u3067\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u56f35\u306f\u5b66\u7fd2\u6642\u306e\u30b5\u30a4\u30ba\u3068\u306f\u7570\u306a\u308b2\u3064\u306e\u30b5\u30a4\u30ba\uff08Inf Res\uff09\uff08\\(768\\times 768\\)\u3068\\(1024\\times 2048\\)\uff09\u306e\u753b\u50cf\u3092\u5165\u529b\u3068\u3057\u305f\u969b\u306e\u4e88\u6e2c\u7cbe\u5ea6\u3092\u63b2\u8f09\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>PE\u306f\u5165\u529b\u30b5\u30a4\u30ba\u306b\u5408\u308f\u305b\u3066\u4f38\u7e2e\u3055\u305b\u3066\u7528\u3044\u3066\u3044\u307e\u3059\u3002Mix-FFN\u3092\u7528\u3044\u305f\u65b9\u304c\u7cbe\u5ea6\u304c\u9ad8\u304f\u3001PE\u3088\u308a\u3082\u4e88\u6e2c\u306e\u6c4e\u5316\u6027\u80fd\u304c\u9ad8\u3044\u3053\u3068\u304c\u5206\u304b\u308a\u307e\u3059\u3002<\/p>\n<p><img decoding=\"async\" class=\"size-full wp-image-17061 aligncenter\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/pe.png\" alt=\"\" width=\"247\" height=\"131\" \/><\/p>\n<p style=\"text-align: center;\">\u56f35\uff1aPE\u3068MixFNN\u306e\u7cbe\u5ea6\u5dee\uff08[1]\u3088\u308a\u629c\u7c8b\uff09<\/p>\n<p>\u6b21\u306b\u3001Decoder\u306e\u7c21\u7565\u5316\u306e\u90e8\u5206\u3092\u898b\u3066\u307f\u307e\u3059\u3002Decoder\u306e\u7c21\u7565\u5316\u306fTransformer\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u306fEncoder\u306e\u51fa\u529b\u304c\u65e2\u306b\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u53d6\u308a\u5165\u308c\u3066\u3044\u308b\u305f\u3081\u8907\u96d1\u306aDecoder\u3092\u5fc5\u8981\u3068\u3057\u306a\u3044\u3068\u3044\u3046\u4e88\u60f3\u306e\u3082\u3068\u884c\u308f\u308c\u307e\u3057\u305f\u3002<\/p>\n<p>\u305d\u306e\u4e88\u60f3\u3092\u78ba\u304b\u3081\u308b\u305f\u3081\u306bDecoder\u3092\u305d\u306e\u307e\u307e\u306b\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306eEncoder\u3092\u7528\u3044\u305f\u5834\u5408\u306e\u7cbe\u5ea6\u3092\u6bd4\u8f03\u3057\u3066\u304a\u308a\u3001\u305d\u306e\u7d50\u679c\u304c\u56f36\u306b\u306a\u308a\u307e\u3059\u3002\u7573\u8fbc\u5c64\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u306b\u6bd4\u3079\u3066\u7cbe\u5ea6\u304c\u9ad8\u304f\u3001\u4e88\u60f3\u3092\u652f\u6301\u3059\u308b\u7d50\u679c\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u307e\u305f\u56f36\u3067\u306f\\(S1-S4\\)\u306f4\u3064\u3059\u3079\u3066\u306e\u30d6\u30ed\u30c3\u30af\u306e\u7279\u5fb4\u91cf\u3092\u7528\u3044\u305f\u3053\u3068\u3092\u3001\\(S4\\)\u306f\u6700\u7d42\u30d6\u30ed\u30c3\u30af\u306e\u7279\u5fb4\u91cf\u306e\u307f\u3092\u7528\u3044\u305f\u3053\u3068\u3092\u8868\u3057\u307e\u3059\u3002<\/p>\n<p>\u3064\u307e\u308a\u3001Transformer\u3092\u5229\u7528\u3057\u305f\u5834\u54084\u3064\u3059\u3079\u3066\u306e\u30d6\u30ed\u30c3\u30af\u3092\u4f7f\u3063\u305f\u7573\u8fbc\u30d9\u30fc\u30b9\u306e\u30e2\u30c7\u30eb\u3088\u308a\u7cbe\u5ea6\u304c\u51fa\u3066\u3044\u308b\u3053\u3068\u306b\u306a\u308a\u3001\u3044\u304b\u306b\u5927\u57df\u7684\u306a\u60c5\u5831\u3092\u52b9\u7387\u7684\u306b\u53d6\u308a\u5165\u308c\u308b\u3053\u3068\u304c\u91cd\u8981\u304b\u3092\u793a\u3057\u305f\u7d50\u679c\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p><img decoding=\"async\" class=\"alignnone wp-image-17062 aligncenter\" src=\"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/enc.png\" alt=\"\" width=\"300\" height=\"160\" \/><\/p>\n<p style=\"text-align: center;\">\u56f36\uff1a\u30d0\u30c3\u30af\u30dc\u30fc\u30f3\u306e\u5dee\uff08[1]\u3088\u308a\u629c\u7c8b\uff09<\/p>\n<h2>\u307e\u3068\u3081<\/h2>\n<p>\u4eca\u56de\u306fTransformer\u69cb\u9020\u3092\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306b\u52b9\u7387\u7684\u306b\u9069\u7528\u3057\u305f\u30e2\u30c7\u30eb\u306e\u7d39\u4ecb\u3067\u3057\u305f\u3002<\/p>\n<p>Attention\u6a5f\u69cb\u3092\u5c0e\u5165\u3057\u3066\u7cbe\u5ea6\u3092\u5411\u4e0a\u3057\u3064\u3064\u3082\u3001\u7279\u5fb4\u91cf\u306e\u968e\u5c64\u69cb\u9020\u5316\u30fbAttention\u6a5f\u69cb\u306e\u52b9\u7387\u5316\u30fbDecoder\u306e\u7c21\u7d20\u5316\u306b\u3088\u3063\u3066\u8a08\u7b97\u91cf\u3092\u3046\u307e\u304f\u6291\u3048\u305f\u30e2\u30c7\u30eb\u3067\u3059\u3002<\/p>\n<p>Attention\u6a5f\u69cb\u52b9\u7387\u5316\u306e\u305f\u3081\u306e\u5165\u529b\u9577\u65b9\u5411\u306e\u5727\u7e2e\u306fK,V\u306b\u5bfe\u3057\u3066\u306f\u5927\u304d\u306a\u30d1\u30c3\u30c1\u3067\u51e6\u7406\u3059\u308b\u306e\u3068\u5909\u308f\u3089\u306a\u3044\u306e\u3067\u3001\u7a7a\u9593\u89e3\u50cf\u5ea6\u306e\u7d30\u304b\u3055\u306fQ\u306e\u307f\u306b\u3088\u3063\u3066\u62c5\u4fdd\u3055\u308c\u3066\u3044\u308b\u3068\u3044\u3046\u69cb\u9020\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u8a08\u7b97\u306e\u8907\u96d1\u5ea6\u306f\u5165\u529b\u9577\u306e2\u4e57\u3067\u3042\u308b\u3053\u3068\u306b\u306f\u5909\u308f\u3089\u305a\u3001\u672c\u8cea\u7684\u306a\u6539\u5584\u3067\u306f\u306a\u3044\u3068\u3044\u3046\u5370\u8c61\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n<p>\u81ea\u7136\u8a00\u8a9e\u306e\u30e2\u30c7\u30eb\u3067\u69d8\u3005\u306a\u52b9\u7387\u7684\u306aAttention\u6a5f\u69cb\u304c\u63d0\u6848\u3055\u308c\u3066\u3044\u308b\u306e\u3067\u3001\u305d\u308c\u3089\u304c\u753b\u50cf\u7cfb\u306e\u30bf\u30b9\u30af\u306e\u7cbe\u5ea6\u306b\u3069\u306e\u3088\u3046\u306a\u5f71\u97ff\u304c\u3042\u308b\u306e\u304b\u6c17\u306b\u306a\u308b\u3068\u3053\u308d\u3067\u3059\u3002<\/p>\n<p>\u8ad6\u6587 [1]\u3067\u306f\u3088\u308a\u8a73\u7d30\u306a\u5206\u6790\u304c\u306a\u3055\u308c\u3066\u3044\u307e\u3059\u306e\u3067\u3001\u6c17\u306b\u306a\u308b\u65b9\u306f\u305c\u3072\u305d\u3061\u3089\u3092\u898b\u3066\u307f\u3066\u304f\u3060\u3055\u3044\u3002<\/p>\n<h2>\u53c2\u8003\u6587\u732e<\/h2>\n<p>[1]Xie, E., Wang, W., Yu, Z., Anandkumar, A., \u00c1lvarez, J.M., &amp; Luo, P. (2021). SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers.\u00a0<em>ArXiv, abs\/2105.15203<\/em>.<\/p>\n<p>[2]Vaswani, A., Shazeer, N.M., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., &amp; Polosukhin, I. (2017). Attention is All you Need.\u00a0<em>ArXiv, abs\/1706.03762<\/em>.<\/p>\n<p>[3]Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., &amp; Schiele, B. (2016). The Cityscapes Dataset for Semantic Urban Scene Understanding.\u00a0<em>2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)<\/em>, 3213-3223.<\/p>\n<p>[4]Everingham, M., Eslami, S.M., Gool, L.V., Williams, C.K., Winn, J.M., &amp; Zisserman, A. (2014). The Pascal Visual Object Classes Challenge: A Retrospective.\u00a0<em>International Journal of Computer Vision, 111<\/em>, 98-136.<\/p>\n<p>[5]Chen, L., Zhu, Y., Papandreou, G., Schroff, F., &amp; Adam, H. (2018). Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation.\u00a0<em>ArXiv, abs\/1802.02611<\/em>.<\/p>\n<p>[6]Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., &amp; Houlsby, N. (2021). An Image is Worth 16&#215;16 Words: Transformers for Image Recognition at Scale.\u00a0<em>ArXiv, abs\/2010.11929<\/em>.<\/p>\n<p>[7]Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., Fu, Y., Feng, J., Xiang, T., Torr, P.H., &amp; Zhang, L. (2021). Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers.\u00a0<em>2021 IEEE\/CVF Conference on Computer Vision and Pattern Recognition (CVPR)<\/em>, 6877-6886.<\/p>\n<p>[8]Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., &amp; Torralba, A. (2018). Semantic Understanding of Scenes Through the ADE20K Dataset.\u00a0<em>International Journal of Computer Vision, 127<\/em>, 302-321.<\/p>\n<p>[9]Caesar, H., Uijlings, J.R., &amp; Ferrari, V. (2018). COCO-Stuff: Thing and Stuff Classes in Context.\u00a0<em>2018 IEEE\/CVF Conference on Computer Vision and Pattern Recognition<\/em>, 1209-1218.<\/p>\n<div class=\"st-header-flexwrap\" style=\"height: auto; width: 100%; box-sizing: border-box; background-color: #efefef; border-radius: 5px; border: solid #ccc 1px; padding: 20px; margin-bottom: 0px;\">\n<div class=\"st-header-flexbox\">\n<p class=\"st-header-flextitle\" style=\"font-size: 110%; color: #000;\">GIS\u3084AI\u6a5f\u68b0\u5b66\u7fd2\u3092\u4f7f\u3063\u305f\u30b7\u30b9\u30c6\u30e0\u306e\u3054\u76f8\u8ac7\u3092\u627f\u3063\u3066\u3044\u307e\u3059\u3002\u304a\u6c17\u8efd\u306b\u304a\u554f\u3044\u5408\u308f\u305b\u304f\u3060\u3055\u3044\u3002<\/p>\n<p><span class=\"st-mybtn st-mybtn-mini\"><a style=\"background: #01579b; border-radius: 5px; font-weight: normal; color: #fff; box-shadow: 0 3px 0 #01579b;\" href=\"https:\/\/www.informatix.co.jp\/contact-us\/\" target=\"_blank\" rel=\"noopener\">\u76f8\u8ac7\u3059\u308b<i class=\"fa fa-after fa-angle-right st-css-no\" aria-hidden=\"true\"><\/i><\/a><\/span><\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u3053\u3093\u306b\u3061\u306f\u3001\u30a4\u30f3\u30d5\u30a9\u30de\u30c6\u30a3\u30af\u30b9\u3067AI\u30fb\u6a5f\u68b0\u5b66\u7fd2\u696d\u52d9\u3092\u62c5\u5f53\u3057\u3066\u3044\u308b\u5927\u6a4b\u3067\u3059\u3002 \u4eca\u56de\u306f\u753b\u50cf\u9818\u57df\u5206\u5272\uff08\u30bb\u30de\u30f3\u30c6\u30a3\u30c3\u30af\u30bb\u30b0\u30e1\u30f3\u30c6\u30fc\u30b7\u30e7\u30f3\uff09\u6280\u8853\u306e\u7d39\u4ecb\u3092\u3057\u305f\u3044\u3068\u601d\u3044\u307e\u3059\u3002\u3053\u3053\u3067\u7d39\u4ecb\u3059\u308b\u8ad6\u6587 [1]\u306f\u3001\u8fd1\u5e74\u767a\u5c55\u8457 &#8230; <\/p>\n","protected":false},"author":11,"featured_media":17051,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[589,4],"tags":[732],"class_list":["post-16949","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-4","tag-transformer"],"jetpack_featured_media_url":"https:\/\/club.informatix.co.jp\/wp-content\/uploads\/2022\/03\/model.png","_links":{"self":[{"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/16949","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=16949"}],"version-history":[{"count":90,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/16949\/revisions"}],"predecessor-version":[{"id":17264,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/posts\/16949\/revisions\/17264"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=\/wp\/v2\/media\/17051"}],"wp:attachment":[{"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=16949"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=16949"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/club.informatix.co.jp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=16949"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}