YOLOã·ãªã¼ãºã®2022å¹´ææ°çãYOLOv7ãã«ã¤ãã¦ãç°å¢æ§ç¯ããå¦ç¿ã®æ¹æ³ã¾ã§ã¾ã¨ãã¾ãã YOLOv7ã¯2022å¹´7æã«å ¬éãããææ°ãã¼ã¸ã§ã³ã§ãããé度ã¨ç²¾åº¦ã®é¢ã§éçãæ¼ãåºãã¦ãã¾ãã 第6åç®ã¯YOLOv7ã«ãã姿å¢æ¨å®ï¼Human Pose Estimationï¼ãç´¹ä»ãã¾ãã Google colabã使ç¨ãã¦ç°¡åã«ææ°ã®ç©ä½æ¤åºã¢ãã«ãå®è£ ãããã¨ãã§ãã¾ãã®ã§ããã²æå¾ã¾ã§ã覧ãã ããã
1.ã¯ããã« é常ã姿å¢æ¨å®ãªã©ã®ã¿ã¹ã¯ã§ã¯ãã¼ããããã使ã£ãå帰ã®ææ³ã使ãã¾ãããä»åãç´¹ä»ããã®ã¯ãã¼ããããã使ããã«å§¿å¢æ¨å®ãè¡ãKapaoã¨ããæè¡ã§ãã ï¼ãã®è«æã¯ã2021.11ã«æåºããã¾ããã 2.Kapaoã¨ã¯ï¼ é常ã姿å¢æ¨å®ãªã©ã®ã¿ã¹ã¯ã§ã¯ãã¼ããããã使ã£ãå帰ã®ã¢ããã¼ããåãã¾ãããçæã¨å¾å¦çã«å¤§éã®è¨ç®å¦çãå¿ è¦ã§ãã Kapaoï¼Keypoints and Poses as Objectsï¼ã¯ããã£ã¨å¦çå¹çãä¸ããããã«ãç»åãç´°ããªã°ãªããã«åå²ãã¦ã人éã®ãã¼ãºãªãã¸ã§ã¯ãã¨ãã¼ãã¤ã³ããªãã¸ã§ã¯ããåæã«æ¤åºã»èåã姿å¢æ¨å®ãè¡ãã¾ãã ä¸è¨ã¯ãKapaoã®ãããã¯ã¼ã¯ã®æ¦è¦ã§ããå ¥åç»åãæ·±ãç³ã¿è¾¼ã¿ãããã¯ã¼ã¯ã§ãããã³ã°ãããã¼ãºãªãã¸ã§ã¯ãã¨ãã¼ãã¤ã³ããªãã¸ã§ã¯ããããããæ¤åºããå¾ããã®ï¼ã¤ã®æ å ±ãèåãçµæãå¾ã¦ãã¾ãã ã§ã¯
ã¯ããã« ããã«ã¡ã¯ã(æ ª) æ¥ç«è£½ä½æ Lumada Data Science Lab. ã®æ£®ç°ã§ãã Lumada Data Science Lab. ã§ã¯ãæ å解ææè¡ã»æ å解æã½ãªã¥ã¼ã·ã§ã³ã®ç 究éçºãè¡ã£ã¦ãã¾ãã ãã®è¨äºã§ã¯ãè¡é ã«ããé²ç¯ã«ã¡ã©ã®æ åã«é©ç¨ããéã«éè¦ã¨ãªããæ¤ç¥ãµã¤ãºãã®è¦³ç¹ã§ã人ç©éª¨æ ¼æ¨å®OSSã®è©ä¾¡ãè¡ãã¾ãã 0. å¿ããæ¹ã¸ã®ã¾ã¨ã 大åã·ã§ããã³ã°ã¢ã¼ã«ãã¹ã¿ã¸ã¢ã ãé§ ã空港ãªã©ã®é²ç¯ã«ã¡ã©ã¯ãåºç¯å²ãå°ãªãå°æ°ã§æ®å½±ãããã®ã§ç»è§ãåºè§ã«ãªããã¡ åºè§ã§æ®å½±ããç»åã¯äººç©ã¯å°ããåãã®ã§ãå°ããåã人ç©ã®éª¨æ ¼ãæ¤åºã§ãããã¨ãéè¦ ä»åã¯éª¨æ ¼æ¨å®AIã®OSSã¨ãã¦ãopenpifpafã¨tf-pose-estimationãããã¯ã¢ããããæå°æ¤ç¥å¯è½ãµã¤ãºã®ææ¨ã§è©ä¾¡ openpifpafã®æ¹ãæå°æ¤ç¥å¯è½ãµã¤ãºãå°ãããã¨ãç¢ºèª 1. é²ç¯
R&D ãã¼ã ã®éååè¦å¯ãã¨å¾³ç°(@dakuton)ã§ãã 7æã®Edge TPUã¢ãããã¼ãã«ãããã¦ã7/30ã«PoseNetã®Edge TPUçãå ¬éããã¦ãã(ä½è Twitteråç»ã¯ãã¡ã)ãããä»å¤åã®ã¤ã³ã¿ã¼ã³ã¨ãã¦ãã£ã¬ã³ã¸ãã¦ãããã¾ããã Edge TPUã¢ãããã¼ãå 容ã«ã¤ãã¦ã¯éå»è¨äºãåç §ãã ããã tech-blog.optim.co.jp PoseNetã¨ã¯ï¼ TensorFlow.js(TensorFlowã®JavaScriptç)ã§åä½ãã姿å¢æ¨å®ã¢ãã«ã§ããWebãã©ã¦ã¶ã§äººã®å§¿å¢ããªã¢ã«ã¿ã¤ã ã«æ¤åºã§ãã(15FPS以ä¸)ã®ãç¹å¾´ã§ãã åèè¨äº: Real-time Human Pose Estimation in the Browser with TensorFlow.js | by TensorFlow | TensorFlow | Mediu
Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields arXiv:https://arxiv.org/abs/1611.08050 æ¦è¦ CNNã®çµåãã§ç»åå ã®äººç©å§¿å¢ãæ¤ç¥ãéæ NP-hardåé¡ãrelaxationãè¨ãã¦è§£ã top-downã¢ããã¼ãï¼äººãæ¤ç¥âæ¤ç¥ãã人ããããã®å§¿å¢æ¤ç¥)ã§ã¯ãªãï¼ bottom-upã¢ããã¼ãï¼å®æéå¦çå¯è½ï¼ COCO 2016 keypoints challengeã¨MPII Multi Person benchmarkã§å¥½æ績ãåºãã Intro ç»åã«æ ã人ç©ã®å§¿å¢æ¨å®ã®é£ãã 人ã®æ°ï¼ã¹ã±ã¼ã«ï¼ä½ç½®ãä¸å®ã§ããã㨠人å士ã®æ¥è§¦ï¼ãªã¯ã«ã¼ã¸ã§ã³ã«ããè¤æ°äººã®å¹²æ¸ 人æ°ã«ãã£ã¦è¨ç®éã大ãããªãã㨠å¾æ¥ææ³ã¯äººã®æ¤ç¥ãè¡ãï¼ãã®å¾ã§å人ã®å§¿å¢æ¨å®ã
çªç¶ã§ãããã¢ãã¡ã¨ãã²ã¼ã ã®ãã£ã©ãèªç±ã«åãããã楽ããã¨æãã¾ãããï¼ ãã¼ãã£ã«YouTuberãããããã§ãããç§ã¯ã·ãããã好ãã§ãããã以å¤ã«ãèªç±èªå¨ã«ãã£ã©ãåããããâ¦â¦ã¨æãã¨å¤¢ãåºããããã§ãã è²ã ãªææ³ãããããã§ãã大ã¾ãã«ã¯ ã»ã¢ã¼ã·ã§ã³ãã£ããã£ã¼ã·ã¹ãã ãã¢ã¼ã·ã§ã³ã»ã³ãµã¼ã§ãã£ãã㣠ã»ã«ã¡ã©æ åããããã£ã½ãçæ ã»ããããã大éã®ã¢ã¼ã·ã§ã³ãäºåç»é²ãã¦ãã®ä¸ããåçãâ¦â¦ãªã© æ©æç¨æããæéãè²»ç¨ãèããã¨ãã«ã¡ã©ã使ããã¨ãæ軽ããã§ãã ã«ã¡ã©æ åããã¢ã¼ã·ã§ã³ãçæããææ³ã®ã²ã¨ã¤ã«ã¤ãã¦èª¿ã¹ã¦ã¿ã¾ããã 姿å¢æ¨å® CMUï¼ã«ã¼ãã®ã¼ã¡ãã³å¤§å¦ï¼ã§å ¬éããã [1611.08050] Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields ã¨ããè«æãããã¾ã
åç¼ã«ã¡ã©ã§æ®å½±ããæ åãå ¥åã«ã人ã®å§¿å¢æ¨å®ããçç¨ããè¡£æã¾ã§åæ§ç¯ãããã¼ã«ã¬ã¹ããã©ã¼ãã³ã¹ãã£ããã£æè¡ãçºè¡¨ 2018-05-16 ç¬ããã¯ã¹ã»ãã©ã³ã¯æ å ±ç§å¦ç 究æï¼Max Planck Institute for Informaticsï¼ã¨ã¹ã¤ã¹é£é¦å·¥ç§å¤§å¦ãã¼ã¶ã³ãæ ¡ï¼EPFLï¼ã®ç 究è ãã¯ã1å°ã®åç¼ã«ã¡ã©ã§æ®å½±ããæ åãå ¥åã«ã人éã®ãã¼ãºããçç¨ããè¡£æã¾ã§ã3Dåæ§ç¯ãããã¼ã«ã¬ã¹ããã©ã¼ãã³ã¹ãã£ããã£æè¡ãMonoPerfCapããçºè¡¨ãã¾ããã è«æï¼MonoPerfCap: Human Performance Capture from Monocular Video èè ï¼Weipeng Xu, Avishek Chatterjee, Michael Zollhoefer, Helge Rhodin, Dushyant Mehta, Hans-Peter
ç»åå¦çã®ç 究éçºãææããã¢ã«ãã©ï¼æ±äº¬é½å代ç°åºï¼ã¯5æ10æ¥ãAIï¼äººå·¥ç¥è½ï¼ãæ´»ç¨ããæ åã«æ ã人ãåç©ã®å§¿å¢ãæ¨å®ããæè¡ãçºè¡¨ãããé é¨ãæãè ãèãªã©ã®é¨ä½ãæ¤åºããããããã®åä½ãã姿å¢ãæ¨å®ã§ããã¨ãããåæ¥ããæä¾ããã æ åãã人ä½18ã«æï¼é¼»ã»ç¼ã»è³ã»é¦ã»è©ã»èã»æé¦ã»è °ã»èã»è¶³é¦ï¼ã®ç¹å¾´ç¹ãæ¤åºãã姿å¢ãæ¨å®ãã¦è¡¨ç¤ºã§ãããAIã®è¦ç´ æè¡ã§ãããã£ã¼ãã©ã¼ãã³ã°ï¼æ·±å±¤å¦ç¿ï¼ã使ããã¨ã§ãé«ç²¾åº¦ãªå§¿å¢æ¨å®ãå®ç¾ããã¨ããã å¾æ¥ã®ã¢ã¼ã·ã§ã³ãã£ããã£ã¼æè¡ã¨ã¯ç°ãªãã対象ã«ãã¼ã«ã¼ãè£ çããå¿ è¦ããªããä½è§£å度ãç½é»ç»åãè¤éãªèæ¯ã§ãæ¤åºã§ããã¨ãããè¤æ°äººã®å§¿å¢æ¨å®ãå¯è½ã§ãå社ãå ¬éããåç»ã«ã¯ã15人以ä¸ãåææ¤åºãã¦ããæ§åãåãããã¦ããã ãã£ã³ã©ã³ãã®Top Data Science社ã¨å ±åéçºãããç£è¦ã«ã¡ã©ã®æ åã使ã£ãç°å¸¸æ¤ç¥ï¼æ¨ªããã£ã¦ããã
Adversarial PoseNet: A Structure-aware Convolutional Network for Human Pose Estimation https://arxiv.org/abs/1705.00389 é¢é£ç 究 姿å¢æ¨å®ã§ã¯DCNN(Deep Convolutional Neural Nets)ã§heatmapãå帰ããææ³ãä¸è¬ç ãããã以ä¸ã®å ´åã§ã¯ç¾å®ã§ã¯ããå¾ãªã姿å¢ãåºåãã¦ãã¾ããã¨ããã é¨ä½ã®occlusionã大ããã¨ãï¼ä»¥ä¸ã®ç»ååç §ï¼ èæ¯ã¨é¨ä½ã®é¡ä¼¼åº¦ãé«ãã¨ã ãããé¿ããã«ã¯äººä½ã®é¢ç¯æ§é ã«ã¤ãã¦ã®äºåæ å ±ãå¿ è¦ã ã§ã人ä½ã®å¹¾ä½çå¶ç´ãDCNNã«å ¥ãè¾¼ãã®ã¯é£ãã Adversarial PoseNet 姿å¢æ¨å®ããçµæãã人ä½ã¨ãã¦å°¤ãããããããimplicitã«å¦ç¿ãããããã«ãGANã®æ çµã¿ãå©ç¨ã é常GANã§
ãªãªã¼ã¹ãé害æ å ±ãªã©ã®ãµã¼ãã¹ã®ãç¥ãã
ææ°ã®äººæ°ã¨ã³ããªã¼ã®é ä¿¡
å¦çãå®è¡ä¸ã§ã
j次ã®ããã¯ãã¼ã¯
kåã®ããã¯ãã¼ã¯
lãã¨ã§èªã
eã³ã¡ã³ãä¸è¦§ãéã
oãã¼ã¸ãéã
{{#tags}}- {{label}}
{{/tags}}