- Learning from Simulated and Unsupervised Images through Adversarial Training
- https://arxiv.org/abs/1612.07828
æ¦è¦
å¦ç¿ã®ããã®ç»åããªãå ´åãä¸è¶³ãã¦ããå ´åï¼ã·ãã¥ã¬ã¼ã¿çã使ã人工çã«ç»åãçæãããã¨ãããï¼ ãããã·ãã¥ã¬ã¼ã¿ã§çæããç»åï¼ä»¥ä¸ã·ãã¥ã¬ã¼ã¿ç»åã¨å¼ç§°ï¼ã¯ï¼ç¾å®ã®ç»åã«ç¾å®æã®è¦³ç¹ã§åã°ãªãï¼ ãã®ããã·ãã¥ã¬ã¼ã¿ç»åã«éå¦ç¿ãã¦ãã¾ãã¨ï¼å®éã®ç»åã«å¯¾ãã¦å¹æãçºæ®ããªããã¨ãããï¼
ã ããã¨è¨ã£ã¦ã·ãã¥ã¬ã¼ã¿ç»åã®è³ªãããããåä¸ãããã®ã¯å¤§å¤ï¼ ããã§ã©ãã«ãªãã®ç¾å®ç»åãè¨ç·´ç»åã¨ãã¦ä½¿ãï¼GANã®æ çµã¿ã§ã·ãã¥ã¬ã¼ã¿ç»åãæ´ç·´ãããï¼ ãã®éï¼ã·ãã¥ã¬ã¼ã¿ç»åã«ä»éããannotationã®æ å ±ã¯ä¿æããããã«ï¼ 大ããªæ¹å¤ã«ã¯ããã«ãã£ãããã(self-regularization loss)ãªã©å·¥å¤«ãæ½ãï¼
SimGAN
SimGANã®ç®çã¯ç¾å®ã®ç»å ã使ã£ã¦ï¼ ã·ãã¥ã¬ã¼ã¿ç»åã ãæ´ç·´ããã¦ç¾å®æãä»ä¸ããå¤æ ãå¦ç¿ããããã¨ï¼
adversarial loss with self-regularization
ãã®éã®è¦ä»¶ã¨ãã¦ï¼ã·ãã¥ã¬ã¼ã¿çæããéã®annotationæ å ±ã¯ä¿æãããã¨ãããã®ãããï¼
ä¾ãã°ï¼è¦ç·æ¨å®ã®ããã®ç»åãã¼ã¿ãæ¡å¼µããå ´åã«ã¯ï¼ ã·ãã¥ã¬ã¼ã¿ã§ç»åçæããéã«è¦ç·æ¹åã®annotationã¯ä¸ãã¦ãããï¼ ãã®annotationããæ å ±ã¯ ã§å¤æããå¾ãä¿æãã¦ããªãã¨ï¼æ師ãã¼ã¿ã¨ãã¦è¦ç·æ¨å®ã«ä½¿ããªãï¼
ããã§é常ã®GANã®Lossã«å ãã¦ï¼ self-regularization lossãå ãããã¨ãææ¡ï¼ self-regularization lossã¯ï¼ãã¯ã»ã«ãã¨ã®å¤æåå¾ã®L1è·é¢ã¨ããï¼
local adversarial loss
SimGANã®discriminatorã®å¦ç¿ã®éï¼ ããç»åç¹å¾´éãé度ã«è©ä¾¡ãã¦ãã¾ãããã«ãªãã¨ï¼ã´ããçæãããã¨ã¨ãªãï¼
ç»åãå±æé åã«åå²ããã¨ãã®åé åãã¨discriminatorã«åºå¥ãããï¼ åå±æé åã«ããã¦ãSimGANã§å¤æããç»åã«ç¾å®æãä»ä¸ããã¦ããç¶æ ã§ã¯ï¼ç¾å®ã®ç»åã®å±æé åã¨åºå¥ãã¤ããªãã¯ãï¼
ããã¤ãã®ç¹å¾´éã«ããè©ä¾¡ãçµã¿åãããã®ã§ï¼ããï¼ã¤ã®ç¹å¾´éã¸ã®ä¾åãå°ãããªãï¼ï¼ããã¯èªä¿¡ããªãï¼
æ¤è¨¼
è¦ç·æ¨å®ã¿ã¹ã¯
è¦ç·æ¨å®ã®å¦ç¿ã«ä½¿ããã¼ã¿ã»ããã¯è³ªãä½ãï¼ ããã§SimGANã§annotationä»ããã¼ã¿ã大éçæãã¦å¦ç¿ãããã¨ããï¼ sotaéæ
ãã¼ã¿ã»ããã«å¯¾ããuser studyãè¡ã£ãï¼ 50åã®ç¾å®ã®ãã¼ã¿ã¨50åã®SimGANã«ããçæãã¼ã¿ãrandomã«æ示ãã¦ã©ã¡ããæ¬ç©ãçãããããæ£ççã¯51.7%ã ã£ãï¼
è·é¢ç»åã«ããæ姿å¢æ¨å®ã¿ã¹ã¯
é常ï¼ç¾å®ã®è·é¢ç»åã«ã¯ãã¤ãºãå ¥ã£ã¦ããï¼ ãã®ãã¤ãºãã·ãã¥ã¬ã¼ã¿ç»åã«ã¯ãªãã®ã§ï¼å¤æãã¦ãã¤ãºãåç¾ï¼