{"id":15416,"date":"2023-11-01T10:24:33","date_gmt":"2023-11-01T02:24:33","guid":{"rendered":"http:\/\/139.9.1.231\/?p=15416"},"modified":"2023-11-01T10:25:23","modified_gmt":"2023-11-01T02:25:23","slug":"eg3d-efficient-geometry-aware-3d-generative-adversarial-networks","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/11\/01\/eg3d-efficient-geometry-aware-3d-generative-adversarial-networks\/","title":{"rendered":"EG3D: Efficient Geometry-aware 3D Generative Adversarial Networks"},"content":{"rendered":"\n<p class=\"has-bright-blue-background-color has-background\">                                       <a href=\"https:\/\/nvlabs.github.io\/eg3d\/\">https:\/\/nvlabs.github.io\/eg3d\/<\/a><\/p>\n\n\n\n<p class=\"has-yellow-background-color has-background\">                 <strong>     \u5c06\u4e09\u7ef4\u5750\u6807\u5bf9\u5e94\u7684\u4f53\u7d20\u7279\u5f81\u5b9a\u4e49\u4e3a\u4e09\u4e2a\u6b63\u4ea4\u6295\u5f71\u5e73\u9762\u7684\u7279\u5f81   <\/strong>    <\/p>\n\n\n\n<p>\u76ee\u524d\u76843D GAN\u8981\u4e48\u8fc7\u4e8e\u8ba1\u7b97\u5bc6\u96c6\u578b\uff0c\u8981\u4e48\u7f3a\u5c11\u591a\u89c6\u56fe\u4e00\u81f4\u6027\uff0c\u8be5\u65b9\u6cd5\u52a0\u5f3a\u4e86\u8ba1\u7b97\u6548\u7387\u5e76\u4e14\u63d0\u5347\u4e86\u91cd\u5efa\u8d28\u91cf\u3002\u4f7f\u7528\u4e86\u663e\u5f0f-\u9690\u5f0f\u7ed3\u6784\uff0c\u4e0d\u4ec5\u751f\u6210\u591a\u89c6\u89d2\u4e00\u81f4\u6027\u56fe\u7247\uff0c\u8fd8\u80fd\u751f\u6210\u9ad8\u8d28\u91cf3D\u51e0\u4f55\u3002\u901a\u8fc7\u89e3\u8026feature generation\u548cneural rendering\uff0c\u8be5\u67b6\u6784\u5c31\u53ef\u4ee5\u7528\u4e0aSOTA\u76842D CNN\u751f\u6210\u5668\u6bd4\u5982styleGAN2\u3002<\/p>\n\n\n\n<p>      \u4f7f\u7528\u5355\u89c6\u89d22D\u56fe\u7247\u96c6\uff0c\u65e0\u76d1\u7763\u5730\u751f\u6210\u9ad8\u8d28\u91cf\u4e14\u89c6\u89d2\u4e00\u81f4\u6027\u5f3a\u76843D\u6a21\u578b\uff0c\u4e00\u76f4\u4ee5\u6765\u90fd\u662f\u4e00\u4e2a\u6311\u6218\u3002\u73b0\u5b58\u76843D GANs\u8981\u4e0d\u8ba1\u7b97\u91cf\u5de8\u5927\uff0c\u8981\u4e0d\u65e0\u6cd5\u4fdd\u8bc13D-consistent\u3002\u524d\u8005\u9650\u5236\u4e86\u751f\u6210\u56fe\u7247\u7684\u8d28\u91cf\uff0c\u540e\u8005\u65e0\u6cd5\u89e3\u51b3\u89c6\u89d2\u4e00\u81f4\u6027\u7684\u95ee\u9898\u3002\u8fd9\u7bc7\u5de5\u4f5c\u63d0\u51fa\u7684\u65b0\u7f51\u7edc\u67b6\u6784\uff0c\u80fd\u53c8\u5feb\u53c8\u597d\u5730\u751f\u62103D geometry\u3002<\/p>\n\n\n\n<p>       \u8fd9\u7bc7\u5de5\u4f5c\u63d0\u51fa\u4e86\u4e24\u4e2a\u65b9\u6cd5\u3002\u9996\u5148\uff0c\u4f5c\u8005\u7528\u663e\u9690\u6df7\u5408\u7684\u65b9\u6cd5\uff0c\u63d0\u9ad8\u4e86\u65f6\u7a7a\u6548\u7387\uff0c\u5e76\u6709\u8f83\u9ad8\u7684\u8d28\u91cf\u3002\u7b2c\u4e8c\uff0c\u63d0\u51fa\u4e86dual-discrimination\u7b56\u7565\uff0c\u4fdd\u8bc1\u4e86\u591a\u89c6\u89d2\u4e00\u81f4\u6027\u3002\u540c\u65f6\uff0c\u8fd8\u5f15\u5165\u4e86pose-based conditioning to the generator\uff0c\u53ef\u4ee5\u89e3\u8026pose\u76f8\u5173\u7684\u53c2\u6570\uff0c\u4fdd\u8bc1\u4e86\u8f93\u51fa\u7684\u89c6\u89d2\u4e00\u81f4\u6027\uff0c\u540c\u65f6\u5fe0\u5b9e\u5730\u91cd\u5efa\u6570\u636e\u96c6\u9690\u542b\u7684pose-correlated\u53c2\u6570\u3002<\/p>\n\n\n\n<p>     \u540c\u65f6\uff0c\u8fd9\u4e2a\u6846\u67b6\u80fd\u89e3\u8026\u7279\u5f81\u751f\u6210\u548c\u795e\u7ecf\u6e32\u67d3\uff0c\u4ece\u800c\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528SOTA\u76842D GANs\uff0c\u6bd4\u5982StyleGAN2\u3002<\/p>\n\n\n\n<h3>contribution<\/h3>\n\n\n\n<ul><li>\u5f15\u5165\u4e00\u4e2a\u57fa\u4e8e\u4e09\u5e73\u9762\u76843D GAN\u67b6\u6784\uff0c\u8ba1\u7b97\u6548\u7387\u9ad8\u800c\u4e14\u6548\u679c\u8d28\u91cf\u597d<\/li><li>\u63d0\u51fa\u4e00\u4e2a3D GAN\u8bad\u7ec3\u7b56\u7565\uff0c\u901a\u8fc7dual discrimination\u548cgenerator pose conditioning\u52a0\u5f3a\u591a\u89c6\u89d2\u4e00\u81f4\u6027\uff0c\u5efa\u6a21\u51fa\u4f4d\u7f6e\u76f8\u5173\u7684\u5c5e\u6027\u5206\u5e03\uff08\u6bd4\u5982\u8868\u60c5\u7b49\uff09<\/li><li>\u5728FFHQ\u548cAFHQ\u4e0a\u6709\u6700\u4f73\u7684\u975e\u6761\u4ef63D\u611f\u77e5\u89c6\u56fe\u5408\u6210\u7ed3\u679c\uff0c\u751f\u6210\u9ad8\u8d28\u91cf3D\u51e0\u4f55<\/li><\/ul>\n\n\n\n<h2 id=\"tri-plane-hybrid-3d-representation\">Tri-Plane Hybrid 3D Representation<\/h2>\n\n\n\n<p>\u6211\u4eec\u9700\u8981\u4e00\u79cd\u9ad8\u6548\u4e14\u8868\u8fbe\u529b\u5f3a\u76843D\u8868\u793a\u65b9\u6cd5\uff0c\u6765\u8bad\u7ec3\u9ad8\u5206\u8fa8\u7387\u7684GAN\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u4ee5\u5355\u573a\u666f\u8fc7\u62df\u5408(SSO)\u6765\u8bc1\u660e\u4e09\u5e73\u9762\u8868\u793a\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n\n\n\n<p>\u6bcf\u4e2a\u5e73\u9762\u90fd\u662fN\u00d7N\u00d7C\u7684\uff0c\u5176\u4e2dC\u662f\u901a\u9053\u6570\u3002<\/p>\n\n\n\n<p>\u6bcf\u6b21\u67e5\u8be2\u4e00\u4e2a3D\u5750\u6807x\u2208R3\uff0c\u5c06\u5176\u6295\u5f71\u81f3\u6bcf\u4e2a\u5e73\u9762\u4e0a\uff0c\u7528\u53cc\u7ebf\u6027\u63d2\u503c\u5f97\u52303\u4e2a\u7279\u5f81\u5411\u91cfFxy,Fxz,Fyz<\/p>\n\n\n\n<p>\u5c06\u8fd93\u4e2a\u7279\u5f81\u5411\u91cf\u7d2f\u52a0\u540e\uff0c\u901a\u8fc7\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684decoder\uff0c\u4e5f\u5c31\u662f\u4e00\u4e2a\u5c0f\u578bMLP\uff0c\u8f93\u51faRGB\u548cDensity<\/p>\n\n\n\n<p>\u518d\u7528volume rendering\u5f97\u5230\u6700\u7ec8\u56fe\u50cf<\/p>\n\n\n\n<p>\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\uff0cdecoder\u89c4\u6a21\u5f88\u5c0f\uff0c\u8d4b\u4e88\u4e86\u663e\u5f0f\u8868\u793a\u66f4\u5f3a\u7684\u8868\u8fbe\u80fd\u529b\uff0c\u5e76\u51cf\u5c0f\u4e86\u8ba1\u7b97\u538b\u529b\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image.png\" alt=\"\" class=\"wp-image-15430\" width=\"366\" height=\"316\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image.png 874w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-300x259.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-768x664.png 768w\" sizes=\"(max-width: 366px) 100vw, 366px\" \/><\/figure><\/div>\n\n\n\n<p>      \u5728\u65b0\u89c6\u89d2\u5408\u6210\u7684\u5b9e\u9a8c\u4e0a\uff0c\u4e09\u5e73\u9762\u7d27\u51d1\u800c\u5bcc\u6709\u8868\u8fbe\u529b\uff0c\u4ee5\u66f4\u4f4e\u7684\u8ba1\u7b97\u6210\u672c\uff0c\u5f97\u5230\u4e86\u66f4\u597d\u7684\u8868\u73b0\uff0c\u4e09\u5e73\u9762\u7684\u65f6\u7a7a\u6210\u672c\u662fO(N2)\u7684\uff0c\u800cvoxel\u662fO(N3)\u7684\uff0c\u6700\u91cd\u8981\u7684\u662f\uff0c\u75282D GANs\u751f\u6210planes\uff0c\u5c31\u80fd\u5f97\u52303D\u8868\u793a\u3002\u5bf9\u6bd4NERF\uff0c\u901a\u8fc7\u663e\u5f0f\u7684\u6295\u5f71\u964d\u4f4e\u4e86\u8ba1\u7b97\u590d\u6742\u5ea6\u540c\u65f6\u6ca1\u6709\u51cf\u5c11\u8868\u8fbe\u6027\u80fd\u3002\u505a\u4e86\u4e2a\u5bf9\u6bd4\u5b9e\u9a8c\uff0cbaseline\u662fmip-nerf\u548cvoxel grid\uff0c\u8fd9\u91cc\u7684tri-plane\u5b9e\u9a8c\u4e2d\u7684MLP\u7528\u4e86\u5085\u91cc\u53f6feature\u7f16\u7801\u3002\u5728\u540c\u6837\u5730\u5185\u5b58\u6d88\u8017\u4e0b\u8fd0\u7b97\u66f4\u5feb\uff0c\u5728\u540c\u6837\u5730\u7ed3\u6784\u4e0b\u901f\u5ea6\u5feb\u4e14\u5185\u5b58\u6d88\u8017\u5c11\u3002<\/p>\n\n\n\n<div class=\"wp-block-image\"><figure class=\"aligncenter size-full is-resized\"><img loading=\"lazy\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-1.png\" alt=\"\" class=\"wp-image-15439\" width=\"388\" height=\"318\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-1.png 802w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-1-300x247.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-1-768x632.png 768w\" sizes=\"(max-width: 388px) 100vw, 388px\" \/><\/figure><\/div>\n\n\n\n<h2>Pipeline<\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" width=\"1024\" height=\"380\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2-1024x380.png\" alt=\"\" class=\"wp-image-15443\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2-1024x380.png 1024w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2-300x111.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2-768x285.png 768w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2-1536x570.png 1536w, http:\/\/139.9.1.231\/wp-content\/uploads\/2023\/11\/image-2.png 1726w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h3 id=\"cnn-generator-backbone--rendering\">CNN Generator Backbone &amp; Rendering<\/h3>\n\n\n\n<p>\u4e09\u5e73\u9762\u7684\u7279\u5f81\uff0c\u662f\u7531StyleGANA\u751f\u6210\u7684\uff0c\u540c\u65f6Latent Code\u548c\u76f8\u673a\u53c2\u6570\u4f1a\u8f93\u5165Mapping Network\uff0c\u751f\u6210\u4e00\u4e2aIntermediate Latent Code<\/p>\n\n\n\n<p>StyleGAN2\u88ab\u4fee\u6539\u540e\uff0c\u8f93\u51fa256\u00d7256\u00d796256\u00d7256\u00d796\u7684\u7279\u5f81\u56fe\uff0c\u4e4b\u540e\u88abreshape\u621032\u901a\u9053\u7684\u5e73\u9762<\/p>\n\n\n\n<p>\u63a5\u7740\u4ece\u4e09\u5e73\u9762\u91c7\u6837\uff0c\u7d2f\u52a0\u540e\uff0c\u901a\u8fc7\u8f7b\u91cf\u7ea7decoder\uff0c\u751f\u6210density\u548c32\u901a\u9053\u7684\u7279\u5f81\uff0c\u7136\u540e\u7531neural volume renderer\u751f\u62102D\u7279\u5f81\u56fe\uff08\u800c\u975eRGB\u56fe\uff09<\/p>\n\n\n\n<h3 id=\"super-resolution\">Super Resolution<\/h3>\n\n\n\n<p>\u4e09\u5e73\u9762\u4ecd\u4e0d\u8db3\u4ee5\u76f4\u63a5\u751f\u6210\u9ad8\u5206\u8fa8\u7387\u56fe\uff0c\u56e0\u6b64\u6dfb\u52a0\u4e86\u8d85\u5206\u6a21\u5757<\/p>\n\n\n\n<p>\u4f7f\u7528\u4e862\u4e2aStyleGAN2\u7684\u5377\u79ef\u5c42\uff0c\u4e0a\u91c7\u6837\u5e76\u4f18\u531632\u901a\u9053\u7279\u5f81\u56fe\uff0c\u5f97\u5230\u6700\u7ec8\u7684RGB\u56fe\u50cf<\/p>\n\n\n\n<h3 id=\"dual-discrimination\">Dual Discrimination<\/h3>\n\n\n\n<p>\u5bf9StyleGAN2\u7684discrimination\u505a\u4e86\u4e24\u4e2a\u4fee\u6539<\/p>\n\n\n\n<p>\u9996\u5148\uff0c\u6dfb\u52a0Dual Discrimination\u4ee5\u4fdd\u8bc1\u751f\u6210\u56fe\u7247\u7684\u89c6\u89d2\u4e00\u81f4\u6027\uff0c\u5373\u4fdd\u8bc1\u539f\u59cb\u56fe\u7247\uff08\u4f4e\u5206\u8fa8\u7387\u751f\u6210\u7684\uff09\u548c\u8d85\u5206\u540e\u7684\u56fe\u7247\u7684\u4e00\u81f4\u6027\uff0c\u5c06\u4f4e\u5206\u8fa8\u7387\u56fe\u7247\u76f4\u63a5\u53cc\u7ebf\u6027\u4e0a\u91c7\u6837\u540e\uff0c\u548c\u8d85\u5206\u56fe\u7247concat\u5f62\u62106\u901a\u9053\u56fe\u7247\uff0c\u771f\u5b9e\u56fe\u7247\u4e5f\u6a21\u7cca\u540e\u7684\u81ea\u5df1\u62fc\u63a5\uff0c\u4e5f\u5f62\u62106\u901a\u9053\u56fe\u7247\uff0c\u8fdb\u884c\u5224\u522b\u3002<\/p>\n\n\n\n<p>\u8fd9\u6837\u505a\uff0c\u4e0d\u4ec5\u80fdencourage\u6700\u7ec8\u8f93\u51fa\u548c\u771f\u5b9e\u56fe\u7247\u7684\u5206\u5e03\u5339\u914d\uff0c\u4e5f\u8ba9\u795e\u7ecf\u6e32\u67d3\u5668\u5c3d\u53ef\u80fd\u5339\u914d\u4e0b\u91c7\u6837\u7684\u771f\u5b9e\u56fe\u7247\uff0c\u5e76\u8ba9\u8d85\u5206\u56fe\u7247\u548c\u795e\u7ecf\u6e32\u67d3\u4fdd\u6301\u4e00\u81f4\u3002<\/p>\n\n\n\n<p>\u5176\u6b21\uff0c\u4f5c\u8005\u5bf9discriminator\u8f93\u5165\u4e86\u76f8\u673a\u5185\u5916\u53c2\uff0c\u4f5c\u4e3a\u4e00\u4e2aconditioning label\uff0c\u4ece\u800c\u8ba9generator\u5b66\u5230\u6b63\u786e\u76843D\u5148\u9a8c\u3002<\/p>\n\n\n\n<h3 id=\"modeling-pose-correlated-attributes\">Modeling Pose-Correlated Attributes<\/h3>\n\n\n\n<p>\u771f\u5b9e\u4e16\u754c\u6570\u636e\u96c6\u5982FFHQ\uff0c\u76f8\u673a\u59ff\u6001\u4e0e\u5176\u4ed6\u53c2\u6570\uff08\u5982\u8868\u60c5\uff09\u6709\u5173\u8054<\/p>\n\n\n\n<p>\u6bd4\u5982\uff0c\u76f8\u673a\u89d2\u5ea6\u4e0e\u4eba\u662f\u5426\u5fae\u7b11\u662f\u6709\u5173\u7cfb\u7684\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u751f\u6210\u7ed3\u679c\u89c6\u89d2\u4e0d\u4e00\u81f4<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u4e3a\u4e86\u66f4\u597d\u7684\u751f\u6210\u8d28\u91cf\uff0c\u9700\u8981\u5c06\u8fd9\u4e9b\u53c2\u6570\u4e0e\u76f8\u673a\u59ff\u6001\u89e3\u8026<\/p>\n\n\n\n<p>\u8fd9\u7bc7\u5de5\u4f5c\u4f7f\u7528\u4e86<strong>Generator Pose Conditioning<\/strong>\u89e3\u8026pose\u548c\u5176\u4ed6\u53c2\u6570<\/p>\n\n\n\n<p>Mapping Network\u4e0d\u4ec5\u63a5\u53d7Latent Code\uff0c\u8fd8\u63a5\u53d7\u76f8\u673a\u53c2\u6570\u505a\u4e3a\u8f93\u5165<\/p>\n\n\n\n<p>\u7ed9\u4e88backbone\u76f8\u673a\u59ff\u6001\u4f5c\u4e3a\u5148\u9a8c\uff0c\u4ece\u800c\u8ba9\u89c6\u89d2\u53ef\u4ee5\u548c\u751f\u6210\u4ea7\u751f\u8054\u7cfb<\/p>\n\n\n\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0cgenerator\u53ef\u4ee5\u5efa\u6a21\u6570\u636e\u96c6\u4e2d\u9690\u5f0f\u7684pose dependent biases\uff0c\u66f4\u5fe0\u5b9e\u5730\u53cd\u6620\u6570\u636e\u96c6\u7279\u5f81<\/p>\n\n\n\n<p>\u4e3a\u4e86\u907f\u514d\u5728\u6e32\u67d3\u65f6\u56e0\u76f8\u673a\u79fb\u52a8\u4ea7\u751f\u89c6\u89d2\u4e0d\u4e00\u81f4\uff0c\u5728\u6e32\u67d3\u65f6\u4fdd\u6301generator\u8f93\u5165\u7684\u76f8\u673a\u53c2\u6570\u4e0d\u53d8<\/p>\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/nvlabs.github.io\/eg3d\/ \u5c06\u4e09\u7ef4\u5750\u6807\u5bf9\u5e94\u7684\u4f53\u7d20\u7279\u5f81\u5b9a\u4e49\u4e3a\u4e09\u4e2a\u6b63\u4ea4\u6295\u5f71\u5e73\u9762\u7684 &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2023\/11\/01\/eg3d-efficient-geometry-aware-3d-generative-adversarial-networks\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">EG3D: Efficient Geometry-aware 3D Generative Adversarial Networks<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[37,4,35],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15416"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=15416"}],"version-history":[{"count":27,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15416\/revisions"}],"predecessor-version":[{"id":15446,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15416\/revisions\/15446"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=15416"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=15416"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=15416"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}