ÔÚCVPR 2020ÉÏ£¬£¬£¬£¬£¬À´×ÔCMU£¬£¬£¬£¬£¬±±¾©´óѧ£¬£¬£¬£¬£¬Ç廪´óѧ£¬£¬£¬£¬£¬ÉÌÌÀÑо¿ÔºµÄÑо¿ÕßÌá³öÁËÒ»ÖÖÎÞ¼àÊÓµÄÈËÌåÐж¯Ç¨áãÒªÁ졶TransMoMo: Invariance-Driven Unsupervised Video Motion Retargeting¡·£¬£¬£¬£¬£¬ÎªÐж¯ÐòÁÐÊý¾Ý½¨Ä£ÌṩÁËÐÂ˼Ð÷¡£¡£¡£¡£¡£¡£¡£

¸ÃÂÛÎÄÌá³öÁËÒ»ÖÖÐÂÓ±µÄÐж¯Öض¨ÏòÍøÂ磨Motion Retargeting Network£©Éè¼Æ£¬£¬£¬£¬£¬¿ÉÒÔÔÚ¶þάҪº¦µã¿Õ¼äÓÉÎÞ±ê×¢µÄÍøÂçÊý¾Ý¶Ëµ½¶ËµØѵÁ·¡£¡£¡£¡£¡£¡£¡£×÷Õß»ùÓÚÎȹÌÐÔÉè¼ÆÁËеÄËðʧº¯Êý£¬£¬£¬£¬£¬´Ó¶ø¸¶ÓëÍøÂçÎÞ¼àÊӵؽâñîÐж¯ÌØÕ÷ÌåÏÖµÄÄÜÁ¦¡£¡£¡£¡£¡£¡£¡£½«ÉÏÊöÐж¯Öض¨ÏòÍøÂçºÍ»ùÓÚÎȹÌÐÔµÄËðʧº¯ÊýÓ¦ÓÃÓÚÈËÌåÐж¯Ç¨áãʹÃüÖÐ£¬£¬£¬£¬£¬ÔÚ¶¨ÐԺͶ¨Á¿Ö¸±êÉ϶¼Áè¼ÝÁËÔ­ÏȵÄ×î¼ÑÒªÁ죨state-of-the-art£©£¬£¬£¬£¬£¬ÓÈÆäÊÇÔÚÕæʵÌìϵÄÖØ´óÐж¯ÉÏÈ¡µÃÏÔ×ŵÄÓÅÊÆ¡£¡£¡£¡£¡£¡£¡£

ʹÃüÏÈÈÝ

ÈËÌåÐж¯Ç¨á㣨Human Motion Transfer£©Ö¸µÄÊǽ«³õʼÔ˶¯ÊÓƵÖгõʼ¹¤¾ßµÄÐж¯Ç¨á㵽ĿµÄ¹¤¾ßÉÏ£¬£¬£¬£¬£¬ÒÔÌìÉúÄ¿µÄÔ˶¯ÊÓƵ¡£¡£¡£¡£¡£¡£¡£ÀýÈç£¬£¬£¬£¬£¬ÒµÓàϲ»¶ÕßÊÇ·ñÄÜÁ¬Ã¦ÏñרҵÈËÊ¿Ò»Ñùѧ»áÌø²î±ðÆø¸ÅµÄÎ赸£¬£¬£¬£¬£¬Èç̽¸ê£¬£¬£¬£¬£¬°ÅÀÙ£¬£¬£¬£¬£¬Èø¶ûÈøÎèµÈ£¿£¿£¿£¿£¿Ö»¹ÜÔÚÏÖʵÖÐÕâÏÕЩÊDz»¿ÉÄܵÄ£¬£¬£¬£¬£¬µ«ÏÖÔÚÈËÃÇ¿ÉÒÔͨ¹ýÈËÌåÐж¯Ç¨áãÊÖÒÕÔÚÐéÄâÌìÏÂʵÏÖÕâһĿµÄ£¬£¬£¬£¬£¬¼´½«Ô´ÊÓƵÖеÄרҵÎ赸ÑÝÔ±µÄÔ˶¯ÐÅϢǨáãµ½ÒµÓàϲ»¶ÕßµÄÐÎÌåÉÏ£¬£¬£¬£¬£¬²¢äÖȾÌìÉúÄ¿µÄÊÓƵ¡£¡£¡£¡£¡£¡£¡£

ÈËÌåÐж¯Ç¨áãÔÚÅÌËã»úͼÐÎѧ¶¯»­ÖÆ×÷ºÍ·ÂÉúÆøеÈËÁìÓòÓÐÆÕ±éµÄÓ¦ÓÃ£¬£¬£¬£¬£¬ºã¾ÃÒÔÀ´ÓÐÐí¶à»ùÓÚÐж¯²¶»ñ×°±¸µÄÑо¿¡£¡£¡£¡£¡£¡£¡£½üÄêÀ´£¬£¬£¬£¬£¬Ëæ×ÅÒƶ¯ÅÌËãµÄÆÕ¼°ºÍÉî¶ÈѧϰÔÚÅÌËã»úÊÓ¾õÁìÓòµÄÓ¦ÓÃ£¬£¬£¬£¬£¬»ñÈ¡ÈËÌåÐж¯ÐÅÏ¢µÄ±¾Ç®´ó´ó½µµÍ¡£¡£¡£¡£¡£¡£¡£±ðµÄ£¬£¬£¬£¬£¬µÃÒæÓÚÌìÉúʽ¶Ô¿¹ÍøÂ磨GAN£©µÄÑо¿Ï£Íû£¬£¬£¬£¬£¬Í¨¹ýÅÌËã»úºÏ³É¸ß¶ÈÕæʵµÄͼÏñºÍÊÓƵ³ÉΪ¿ÉÄÜ¡£¡£¡£¡£¡£¡£¡£Òò´Ë£¬£¬£¬£¬£¬Éè¼Æ³öÒ»ÖÖºÏÊʵĻùÓÚÊÓƵÊý¾ÝµÄÈËÌåÐж¯Ç¨áãÒªÁì¶ÔÐí¶àÓ¦ÓÃÁìÓò¶¼Ê®·ÖÖ÷Òª£¬£¬£¬£¬£¬ÀýÈçÓ°ÊÓЧ¹û¡¢ÄÚÈÝ´´×÷¡¢ÔÚÏßÓéÀÖµÈ£¬£¬£¬£¬£¬ÕâÒ²ÊDZ¾ÎÄÏëҪ̽ÌÖµÄÖصã¡£¡£¡£¡£¡£¡£¡£

ÄîÍ·/Äѵã

ÒÑÍù´¦Öóͷ£ÈËÌåÐж¯Ç¨áãµÄÎÊÌâÉÏ£¬£¬£¬£¬£¬ÓÐÐí¶àÑо¿Í¨¹ý׼ȷµÄ3DÈËÌå×Ë̬Ԥ¼Æ»òÖØÐÞÀ´ÊµÏÖ¡£¡£¡£¡£¡£¡£¡£ÕâЩҪÁìÍùÍùÐèÒªÖØ´óºÄʱµÄÓÅ»¯Àú³Ì£¬£¬£¬£¬£¬²¢ÇÒÔÚÕæʵ³¡¾°ÖеÄÖØ´óÐж¯ÊÓƵÖÐÌåÏÖ²»¼Ñ¡£¡£¡£¡£¡£¡£¡£ÔÚ±¾ÎÄÖÐ£¬£¬£¬£¬£¬×÷ÕßÖ÷ÒªÉè¼ÆÁËÒ»ÖÖ¿ÉÔÚ¶þάÈËÌå¹Ç÷À¿Õ¼ä¶Ëµ½¶ËѧϰµÄÐж¯Ç¨áã¿ò¼Ü£¬£¬£¬£¬£¬´Ó¶øÈƹýÁËÏÔʽµÄ3D×Ë̬Ԥ¼ÆÀú³Ì¡£¡£¡£¡£¡£¡£¡£Ö»¹Ü×î½üÔÚÌìÉúÄ£×ÓºÍÈËÌåÐж¯ºÏ³ÉÁìÓòÓÐÐí¶àÏ£Íû£¬£¬£¬£¬£¬ÔÚ2D¿Õ¼äѧϰÐж¯Ç¨áãÈÔÈ»ÊÇÒ»¸ö¸»ÓÐÌôÕ½ÐÔµÄʹÃü£¬£¬£¬£¬£¬Ö÷ÒªÓÐÒÔÏÂÄѵ㣺

1. ÓÉÓÚ³õʼÔ˶¯ÊÓƵºÍÄ¿µÄÔ˶¯ÊÓƵ±£´æºÜ´óµÄ½á¹¹ºÍÊӽDzî±ð£¬£¬£¬£¬£¬ºÜÄÑÔÚÏñËؼ¶±ðÉϽ¨ÉèÔ´-Ä¿µÄµÄÓ³Éä¡£¡£¡£¡£¡£¡£¡£ÓÈÆäµ±³õʼ¹¤¾ß×ö³öÖØ´óÐж¯£¨ÈçÌɵ¹£¬£¬£¬£¬£¬µ¹Á¢£¬£¬£¬£¬£¬Ò»Á¬ÐýתµÈ£©£¬£¬£¬£¬£¬»òÕß³õʼ¹¤¾ßºÍÄ¿µÄ¹¤¾ßµÄ½á¹¹²î±ð½ÏÁ¿´ó£¨Èç³ÉÈËÓë¶ùͯµÈ£©£¬£¬£¬£¬£¬¹Å°åµÄÐж¯Ç¨áãÒªÁì׼ȷ¶È½ÏµÍ¡£¡£¡£¡£¡£¡£¡£

2. ºÜÄÑÔÚÕæʵÌìÏÂÖÐÕÒµ½Åä¶ÔµÄÐж¯-½ÇÉ«Êý¾Ý×÷ΪÈËÌåÐж¯Ç¨áãʹÃüµÄÓÐÓüàÊÓÐźÅ¡£¡£¡£¡£¡£¡£¡£

3. ÈËÌåÔ˶¯·ºÆð³öÖØ´óµÄ·ÇÏßÐÔ£¬£¬£¬£¬£¬ºÜÄѽ¨Éè׼ȷµÄÄ£×ӺͲÎÊýÀ´Ãè»æÈËÌåÐж¯Ç¨áãµÄÀú³Ì¡£¡£¡£¡£¡£¡£¡£

ÒªÁìÏÈÈÝ

¿ò¼ÜÉè¼Æ

ΪÁËÓ¦¶ÔµÚÒ»¸öÄѵã£¬£¬£¬£¬£¬×÷Õß²»Ö±½Ó½ÓÄÉ»ùÓÚÏñËؼ¶Ó³ÉäµÄÊÓƵת»»£¨video-to-video translation£©ÒªÁì£¬£¬£¬£¬£¬¶øÊǽ«Ðж¯Ç¨áãÀú³Ì·ÖΪÈý¸ö½×¶Î£¬£¬£¬£¬£¬¼´ÈËÌåÒªº¦µã¼ì²â£¨skeleton extraction£©£¬£¬£¬£¬£¬Ðж¯Öض¨Ïò£¨motion retargeting£©£¬£¬£¬£¬£¬ÒÔ¼°ÊÓƵäÖȾ£¨skeleton-to-video rendering£©¡£¡£¡£¡£¡£¡£¡£Í¨¹ý¶ÔʹÃüµÄÆÊÎö£¬£¬£¬£¬£¬Ö»ÐèÒª¼¯Öнâ¾öÐж¯Öض¨ÏòÕâÒ»ÎÊÌâ£¬£¬£¬£¬£¬´ËÎÊÌâµÄÊäÈëºÍÊä³ö¶¼ÊÇ2DÈËÌåÒªº¦µãÐòÁС£¡£¡£¡£¡£¡£¡£

ÕûÌå¿ò¼ÜʾÒâͼ

ΪÁËÓ¦¶ÔµÚ¶þºÍµÚÈý¸öÄѵã£¬£¬£¬£¬£¬×÷ÕßʹÓÃÈËÌåÔ˶¯Êý¾ÝÖÐÈý¸öά¶ÈµÄÌØÕ÷µÄÎȹÌÐÔ¡£¡£¡£¡£¡£¡£¡£Èý¸öά¶ÈµÄÌØÕ÷¼´Ô˶¯£¨motion£©£¬£¬£¬£¬£¬Ö¸µÄÊÇÉíÌå¸÷²¿·ÖµÄÒƶ¯ÓïÒåÐÅÏ¢£»£»£»£»£»£»½á¹¹£¨structure£©£¬£¬£¬£¬£¬Ö¸µÄÊÇÉíÌåµÄ±ÈÀýÐÅÏ¢£»£»£»£»£»£»Êӽǣ¨view£©£¬£¬£¬£¬£¬Ö¸µÄÊÇÉíÌåºÍÏà»úµÄÏà¶Ô³¯ÏòÐÅÏ¢¡£¡£¡£¡£¡£¡£¡£ÀíÂÛÉÏ£¬£¬£¬£¬£¬Æ¾Ö¤ÕâÈý¸öÐÅÏ¢Äܹ»ÖØÐÞÕûÌåÔ˶¯£¬£¬£¬£¬£¬ÇÒÕâÈý²¿·ÖÐÅÏ¢Ï໥×ÔÁ¦£¬£¬£¬£¬£¬ÈÎÒ»ÐÅÏ¢¶ÔÁíÁ½ÕßµÄÈŶ¯£¨perturbation£©¾ß±¸ÎȹÌÐÔ¡£¡£¡£¡£¡£¡£¡£ÏêϸµØÀ´Ëµ£¬£¬£¬£¬£¬×÷ÕßÒÔΪËüÃÇÓ¦¾ßÓÐÈçÏÂÐÔ×Ó£º

1. Ô˶¯ÐÅÏ¢Ó¦¸Ã¹ØÓڽṹºÍÊÓ½ÇÈŶ¯¾ßÓÐÎȹÌÐÔ¡£¡£¡£¡£¡£¡£¡£

2. ½á¹¹ÐÅÏ¢Ó¦¸ÃÔÚͳһ¶ÎÊÓƵÖйØÓÚʱ¼ä¾ßÓÐÎȹÌÐÔ£¬£¬£¬£¬£¬ÇÒ¶ÔÊÓ½ÇÈŶ¯ÓÐÎȹÌÐÔ¡£¡£¡£¡£¡£¡£¡£

3. ÊÓ½ÇÐÅÏ¢Ó¦¸ÃÔÚͳһ¶ÎÊÓƵÖйØÓÚʱ¼ä¾ßÓÐÎȹÌÐÔ£¬£¬£¬£¬£¬ÇҶԽṹÈŶ¯ÓÐÎȹÌÐÔ¡£¡£¡£¡£¡£¡£¡£

ѵÁ·ÊµÏÖ

ѵÁ·ÊµÏÖÉÏ£¬£¬£¬£¬£¬×÷ÕßʹÓÃ3DÈËÌåµÄÐýת×÷ΪÊÓ½ÇÐÅÏ¢µÄÈŶ¯£»£»£»£»£»£»Ö«ÌåËõ·Å£¨Limb Scaling£©×÷Ϊ½á¹¹ÐÅÏ¢µÄÈŶ¯£»£»£»£»£»£»Ô˶¯ÐÅÏ¢ÎÞÐèÏÔʽÌí¼ÓÈŶ¯£¬£¬£¬£¬£¬ÓÉÓÚÆä×ÔÉí¾ÍËæʱ¼äת±ä¡£¡£¡£¡£¡£¡£¡£»£»£»£»£»£»ùÓÚÕâЩÈŶ¯£¬£¬£¬£¬£¬×÷ÕßÒªÇóÍøÂçÖØбàÂëµÄÌØÕ÷»®·Ö¾ßÓÐÉÏÊöÎȹÌÐÔ£¬£¬£¬£¬£¬¿ÉÒÔÍƵ¼³öһϵÁÐÍêÈ«ÎÞ¼àÊÓµÄËðʧº¯Êý£¬£¬£¬£¬£¬½ø¶øͨ¹ýѵÁ·×Ô±àÂëÆ÷£¨Autoencoder£©½«ÈËÌåÒªº¦µãÐòÁÐÐÅÏ¢½âñîΪÔ˶¯¡¢½á¹¹¡¢ÊÓ½ÇÈý¸öÏ໥Õý½»µÄÐÅÏ¢·ÖÁ¿¡£¡£¡£¡£¡£¡£¡£

ÊÓ½ÇÐÅÏ¢µÄÈŶ¯£¨View Perturbation£©¼°ÖØÐÞ

½á¹¹ÐÅÏ¢µÄÈŶ¯£¨Structural Perturbation£©¼°ÖØÐÞ

ÈçͼËùʾ£¬£¬£¬£¬£¬Ïà¹Øϸ½ÚÇë²Î¿¼ÂÛÎÄ¡£¡£¡£¡£¡£¡£¡£

½èÖú½âñîµÄÌåÏÖ£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔºÜÈÝÒ×µØδÀ´×Ô²î±ðÊÓƵµÄÔ˶¯·ÖÁ¿ºÍ½á¹¹·ÖÁ¿Åä¶ÔºóÔÙ½âÂëµ½Òªº¦µã¿Õ¼ä£¬£¬£¬£¬£¬¾ÍʵÏÖÁËÐж¯Öض¨Ïò¡£¡£¡£¡£¡£¡£¡£±ðµÄ£¬£¬£¬£¬£¬ÎÒÃÇ»¹¿ÉÒÔÏÔʽµØʹÓÃÊÓ½ÇÐÅÏ¢£¬£¬£¬£¬£¬»ñµÃÌìÉúÐж¯ÐòÁÐÔÚ²î±ðÊÓ½ÇϵÄͶӰ¡£¡£¡£¡£¡£¡£¡£ÖµµÃÒ»ÌáµÄÊÇ£¬£¬£¬£¬£¬×÷ÕßÌá³öµÄÐж¯Öض¨ÏòÍøÂçÊÇÒ»¸öÇáÁ¿¼¶¡¢¼´²å¼´ÓõÄÄ£¿£¿£¿£¿£¿é£¬£¬£¬£¬£¬¿ÉÒÔºÍÏÖÓеÄÖÖÖÖÈËÌåÒªº¦µã¼ì²âºÍÊÓƵäÖȾҪÁì´®ÁªÊ¹Óᣡ£¡£¡£¡£¡£¡£

Ч¹ûչʾ

×÷Õß»®·ÖÔÚMixamoºÏ³É½ÇÉ«Êý¾Ý¼¯ºÍ×Ô¼ºÊÕÂÞµÄÕæʵÈËÎïÊý¾Ý¼¯ÉÏ×öÁ˶¨Á¿ºÍ¶¨ÐÔµÄÆÀ²â¡£¡£¡£¡£¡£¡£¡£Ð§¹ûÏÔʾ£¬£¬£¬£¬£¬±¾ÊÂÇéÔÚÐж¯Ç¨áãÎó²î£¨MSE/MAE£©ºÍÌìÉúͼÏñµÄÕæʵ¶È£¨FID/User Preference£©É϶¼±ÈÏÖÔÚµÄSOTAÒªÁìÓÐÏÔÖøÌáÉý£¬£¬£¬£¬£¬Ð§¹û¼ûTable1¡£¡£¡£¡£¡£¡£¡£

ÆäÖÐ£¬£¬£¬£¬£¬¸ÃÊÂÇéÓÃÁËѵÁ·Êý¾ÝÖÐ×îÉÙµÄÐÅÏ¢£¬£¬£¬£¬£¬¼ÈûÓÐÓýÇÉ«-Ðж¯¼äµÄÅä¶ÔÐÅÏ¢£¬£¬£¬£¬£¬Ò²Ã»ÓÐÓÃÊäÈëµÄ3DÐÅÏ¢¡£¡£¡£¡£¡£¡£¡£±ðµÄ£¬£¬£¬£¬£¬×÷ÕßÌá³öµÄÒªÁìʹµÃÔÚí§ÒâÈËÌåÔ˶¯ÊÓƵÉÏѵÁ·³ÉΪ¿ÉÄÜ£¬£¬£¬£¬£¬ÕâÊÇ´ËÇ°µÄÒªÁìûÓв½·¥×öµ½µÄ¡£¡£¡£¡£¡£¡£¡£ÊµÑéЧ¹ûÅú×¢£¬£¬£¬£¬£¬Ê¹ÓÃÍøÉÏÊÕÂÞµÄÕæʵÊÓƵѵÁ·µÄÄ£×Ó£¨Ours wild£©È¡µÃÁË×îºÃµÄÌåÏÖ¡£¡£¡£¡£¡£¡£¡£

¸ÃÒªÁ컹ÔÊÐí±àÂë¿Õ¼äµÄÒ»Á¬²åÖµ¡£¡£¡£¡£¡£¡£¡£ÈçͼËùʾ£¬£¬£¬£¬£¬ºáÏòÊǽṹÐÅÏ¢µÄÌØÕ÷²åÖµ£¬£¬£¬£¬£¬×ÝÏòÊÇÐж¯ÐÅÏ¢µÄÌØÕ÷²åÖµ¡£¡£¡£¡£¡£¡£¡£ÕâÅú×¢ÍøÂçÄܹ»ÀֳɽâñîÕâÁ½¸ö±äÁ¿£¬£¬£¬£¬£¬²¢ÇÒѧϰÁ˾ßÓÐÓÅÒìÂþÑÜÐÔ×ÓµÄÌØÕ÷×ӿռ䡣¡£¡£¡£¡£¡£¡£

±ðµÄ£¬£¬£¬£¬£¬ÓÉÓÚ×Ô¶¯±àÂëÆ÷µÄÊä³öÊÇ3DÑó»ðÈË£¬£¬£¬£¬£¬ÎÒÃÇ¿ÉÒÔí§ÒâµØµ÷½âÖض¨ÏòЧ¹ûµÄÊÓ½ÇÆ«Ïò¡£¡£¡£¡£¡£¡£¡£

ÆÊÎö×ܽá

×ÜÌåÀ´Ëµ£¬£¬£¬£¬£¬×÷ÕßÌá³öµÄÎÞ¼àÊÓÐж¯Ç¨áãÒªÁìÔÚ¶ÔѵÁ·Êý¾ÝµÄÒÀÀµË®Æ½¡¢Ç¨áãµÄ׼ȷ¶È¡¢ÌìÉúЧ¹ûµÄÖÊÁ¿µÈ·½Ãæ¾ù±È֮ǰµÄͬÀàÐÍÒªÁìÓÐÍ»ÆÆ¡£¡£¡£¡£¡£¡£¡£ÕâÖÖÐÔÄÜÉϵÄÌáÉýÖ÷ҪȪԴÓÚÒÔÏÂÁ¢Òìµã£º

• ÔÚ2DÒªº¦µã¿Õ¼äʵÏÖÌØÕ÷½âñî£¬£¬£¬£¬£¬×èÖ¹ÁË3DÒªº¦µãÔ¤¼ÆµÄÎó²î

• ºÍÏÈÇ°µÄÒþʽÎÞ¼àÊÓÒªÁìÏà±È£¬£¬£¬£¬£¬ÏÔʽµÄÎȹÌÐÔÇý¶¯µÄËðʧº¯ÊýÌá¸ßÁËѵÁ·Êý¾ÝЧÂÊ£¨Data efficiency£©

• Ö«ÌåËõ·ÅµÄÉè¼ÆÌáÉýÁËÍøÂç´¦Öóͷ£ÉíÌå±ÈÀý²î±ð½Ï´óÇéÐεÄÄÜÁ¦

• ºÍÏÈÇ°ÔںϳÉÊý¾Ý¼¯ÉϵļàÊÓѧϰҪÁìÏà±È£¬£¬£¬£¬£¬Ê¹ÓÃÍøÂçÊý¾ÝÎÞ¼àÊÓѵÁ·´ó´óÔöÌíÁËѵÁ·Ê±µÄÐж¯¶àÑùÐÔ

ÁíÍâ£¬£¬£¬£¬£¬¸ÃÊÂÇéÉè¼ÆµÄÐж¯Öض¨ÏòÍøÂçʵÏÖÁËÎÞ¼àÊÓÌØÕ÷½âñî£¬£¬£¬£¬£¬¿ÉÄܶÔÈËÌåÔ˶¯½¨Ä£µÄÆäËûÎÊÌâÓÐÒ»¶¨µÄÆô·¢¡£¡£¡£¡£¡£¡£¡£

References

1. Ruben      Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic      networks for unsupervised motion retargetting. In CVPR, 2018.

2. Caroline      Chan, Shiry Ginosar, Tinghui Zhou, and Alexei A. Efros. Everybody dance      now. In ICCV, 2019.

3. Kfir      Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or.      Learning character-agnostic motion for motion retargeting in 2d. ACM      Trans. Graph., 38(4):75:1¨C75:14, 2019.

4. Ting-Chun      Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and      Bryan Catanzaro. Video-to-video synthesis. In NeurIPS, 2018.

5. R?za      Alp G¨¹ler, Natalia Neverova, and Iasonas Kokkinos. Densepose: Dense human      pose estimation in the wild. In CVPR, pages 7297¨C7306, 2018.