基于论述内容，可以得出以下20字中文标题：测试时自适应：自监督对比学习与在线伪标签细化的方法

AB6nicbVDLSgNBEOyNrxhfUY9eBoPgKexKUI8BL4KXiOYByRJmJ7PJkNnZaZXDEvAH/DiQRGvfpE3/8bJ46CJBQ1FVTfdXUEihUHX/XZyK6tr6xv5zcLW9s7uXnH/oGHiVDNeZ7GMdSughkuheB0FSt5KNKdRIHkzGF5N/OYD10bE6h5HCfcj2lciFIyile4eu163WHL7hRkmXhzUoI5at3iV6cXszTiCpmkxrQ9N0E/oxoFk3xc6KSGJ5QNaZ+3LVU04sbPpqeOyYlVeiSMtS2FZKr+nshoZMwoCmxnRHFgFr2J+J/XTjG89DOhkhS5YrNFYSoJxmTyN+kJzRnKkSWUaWFvJWxANWVo0ynYELzFl5dJ46zsnZcrt5VS9eZpFkcejuAYTsGDC6jCNdSgDgz68Ayv8OZI58V5dz5mrTlnHuEh/IHz+QM3U4yx1AB6nicbVDLSgNBEOyNrxhfUY9eBoPgKeyGoB4DXgQvEU0iJEuYnfQmQ2Znl5lZMSwBf8CLB0W8+kXe/Bsnj4MmFjQUVd10dwWJ4Nq47reTW1ldW9/Ibxa2tnd294r7B0dp4phg8UiVvcB1Si4xIbhRuB9opBGgcBWMLyc+K0HVJrH8s6MEvQj2pc85IwaK90+divdYsktu1OQZeLNSQnmqHeLX51ezNIpWGCat323MT4GVWGM4HjQifVmFA2pH1sWyphNrPpqeOyYlVeiSMlS1pyFT9PZHRSOtRFNjOiJqBXvQm4n9eOzXhZ9xmaQGJZstClNBTEwmf5MeV8iMGFlCmeL2VsIGVFmbDoFG4K3+PIyaVbK3lm5elMt1a6fZnHk4QiO4RQ8OIcaXEdGsCgD8/wCm+OcF6cd+dj1pz5hEewh84nz8414zx2AB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆy(c) SHOTAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB8HicbVBNS8NAEJ3Ur1q/qh69BIvgqSlqMeCF71VsK3ShrLZbtqlu5uwOxFK6K/w4kERr/4cb/4bt20O2vpg4PHeDPzwkRwg5737RTW1jc2t4rbpZ3dvf2D8uFR28SpqxFYxHrh5AYJrhiLeQo2EOiGZGhYJ1wfD3zO09MGx6re5wkLJBkqHjEKUErPfZGBLPJtF/rlyte1ZvDXSV+TiqQo9kvf/UGMU0lU0gFMabrewkGdHIqWDTUi81LCF0TIasa6kikpkgmx8dc+sMnCjWNtS6M7V3xMZkcZMZGg7JcGRWfZm4n9eN8XoKsi4SlJki4WRalwMXZn37sDrhlFMbGEUM3trS4dEU0o2oxKNgR/+eV0q5V/Ytq/a5eadzmcRThBE7hHy4hAbcQBNaQEHCM7zCm6OdF+fd+Vi0Fpx85hj+wPn8AeLykH8=ˆy2AB8HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkqMeCF71VsB/ShrLZbtulu0nYnQgh9Fd48aCIV3+ON/+N2zYHbX0w8Hhvhpl5QSyFQdf9dgpr6xubW8Xt0s7u3v5B+fCoZaJEM95kYx0J6CGSxHyJgqUvBNrTlUgeTuY3Mz89hPXRkThA6Yx9xUdhWIoGEUrPfbGFLN02vf65Ypbdecgq8TLSQVyNPrlr94gYoniITJjel6box+RjUKJvm01EsMjymb0BHvWhpSxY2fzQ+ekjOrDMgw0rZCJHP190RGlTGpCmynojg2y95M/M/rJji89jMRxgnykC0WDRNJMCKz78lAaM5QpZQpoW9lbAx1ZShzahkQ/CWX14lrYuqd1mt3dcq9bs8jiKcwCmcgwdXUIdbaEATGCh4hld4c7Tz4rw7H4vWgpPHMfOJ8/4W6Qfg==ˆy1AB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB8HicbVBNS8NAEJ3Ur1q/qh69BIvgqSlqMeCF71VsK3ShrLZbtqlu5uwOxFK6K/w4kERr/4cb/4bt20O2vpg4PHeDPzwkRwg5737RTW1jc2t4rbpZ3dvf2D8uFR28SpqxFYxHrh5AYJrhiLeQo2EOiGZGhYJ1wfD3zO09MGx6re5wkLJBkqHjEKUErPfZGBLPJtF/rlyte1ZvDXSV+TiqQo9kvf/UGMU0lU0gFMabrewkGdHIqWDTUi81LCF0TIasa6kikpkgmx8dc+sMnCjWNtS6M7V3xMZkcZMZGg7JcGRWfZm4n9eN8XoKsi4SlJki4WRalwMXZn37sDrhlFMbGEUM3trS4dEU0o2oxKNgR/+eV0q5V/Ytq/a5eadzmcRThBE7hHy4hAbcQBNaQEHCM7zCm6OdF+fd+Vi0Fpx85hj+wPn8AeLykH8=ˆy2AB8HicbVBNS8NAEJ3Ur1q/qh69LBbBU0mkqMeCF71VsB/ShrLZbtulu0nYnQgh9Fd48aCIV3+ON/+N2zYHbX0w8Hhvhpl5QSyFQdf9dgpr6xubW8Xt0s7u3v5B+fCoZaJEM95kYx0J6CGSxHyJgqUvBNrTlUgeTuY3Mz89hPXRkThA6Yx9xUdhWIoGEUrPfbGFLN02vf65Ypbdecgq8TLSQVyNPrlr94gYoniITJjel6box+RjUKJvm01EsMjymb0BHvWhpSxY2fzQ+ekjOrDMgw0rZCJHP190RGlTGpCmynojg2y95M/M/rJji89jMRxgnykC0WDRNJMCKz78lAaM5QpZQpoW9lbAx1ZShzahkQ/CWX14lrYuqd1mt3dcq9bs8jiKcwCmcgwdXUIdbaEATGCh4hld4c7Tz4rw7H4vWgpPHMfOJ8/4W6Qfg==ˆy1AB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆyAB6HicbVDLTgJBEOzF+IL9ehlIjHxRHYNUY8kXjyCkUcCGzI79MLI7OxmZtZICF/gxYPGePWTvPk3DrAHBSvpFLVne6uIBFcG9f9dnJr6xubW/ntws7u3v5B8fCoqeNUMWywWMSqHVCNgktsG4EthOFNAoEtoLRzcxvPaLSPJb3ZpygH9GB5CFn1Fip/tQrltyOwdZJV5GSpCh1it+dfsxSyOUhgmqdcdzE+NPqDKcCZwWuqnGhLIRHWDHUkj1P5kfuiUnFmlT8JY2ZKGzNXfExMaT2OAtsZUTPUy95M/M/rpCa89idcJqlByRaLwlQE5PZ16TPFTIjxpZQpri9lbAhVZQZm03BhuAtv7xKmhdl7JcqVdK1bsjycwCmcgwdXUIVbqEDGCA8wyu8OQ/Oi/PufCxac042cwx/4Hz+AO3jRU=xAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPBi8cq9gPaUDbTbt0swm7EyGE/gvHhTx6u/x5r9x2+agrQ8GHu/NMDMvSKQw6LrfTmltfWNzq7xd2dnd2z+oHh61TZxqxlslrHuBtRwKRvoUDJu4nmNAok7wST25nfeLaiFg9YpZwP6IjJULBKFqp0x9TzLPpoFpz6+4cZJV4BalBgeag+tUfxiyNuEImqTE9z03Qz6lGwSfVvqp4QlEzriPUsVjbjx8/m5U3JmlSEJY21LIZmrvydyGhmTRYHtjCiOzbI3E/zeimGN34uVJIiV2yxKEwlwZjMfidDoTlDmVlCmRb2VsLGVFOGNqGKDcFbfnmVtC/q3lX98v6y1ngo4ijDCZzCOXhwDQ24gya0gMEnuEV3pzEeXHenY9Fa8kpZo7hD5zPH7l/j+M=ˆy2950对比测试时自适应0Dian Chen * 丰田研究所 Dequan Wang 加州大学伯克利分校 Trevor Darrell 加州大学伯克利分校 Sayna Ebrahimi †0加州大学伯克利分校0摘要0测试时自适应是无监督领域自适应的一种特殊情况，其中在源域上训练的模型必须在没有访问源数据的情况下适应目标域。我们提出了一种新颖的方法，利用自监督对比学习来促进目标特征学习，同时采用在线伪标签方案进行细化，显著去噪伪标签。对比学习任务与伪标签一起应用，构造类似于MoCo的正负样本对，但使用源初始化的编码器，并排除由伪标签指示的相同类别的负样本对。同时，我们在线生成伪标签，并通过在目标特征空间中的最近邻之间进行软投票来细化它们，这是通过维护一个内存队列实现的。我们的方法AdaContrast在主要基准测试中取得了最先进的性能，同时与现有方法相比具有几个可取的特性，包括内存效率、对超参数的不敏感性和更好的模型校准。代码已在https://github.com/DianCh/AdaContrast上发布。01. 引言0深度网络在训练和测试数据遵循相同分布的视觉任务中取得了显著的成功。然而，在面对领域转移时，它们在泛化到未见数据方面的能力会受到影响[42,43]。构建能够适应分布转移的模型是领域自适应的重点，其目标是将知识从标记的源域转移到新的但相关的目标域[2,13, 19, 31, 44, 53]。在这项工作中，我们专注于测试时[50,56]或无源[23, 28,58]领域自适应问题，在自适应到未标记的测试数据时不再有源数据可用。由于测试时自适应（TTA）只需要访问源模型，因此它在数据隐私和传输带宽成为关键问题的实际应用中具有吸引力。0* 作者在加州大学伯克利分校工作时完成的工作。†作者现在在谷歌工作。0(d) AdaContrast (我们的方法)0对比0学习0在线0细化0离线0细化0(b) 伪标签0(a) 仅源域0图1.我们的方法AdaContrast如何利用目标域数据与之前的方法进行对比的示意图。 (a)在没有自适应的情况下，仅源域模型仅在目标数据上进行评估。 (b)使用伪标签，源分类器的预测被用作自训练的伪标签。 (c)现有的伪标签方法SHOT[28]使用离线全局细化来减少嘈杂的伪标签。 (d)在AdaContrast中，我们考虑目标样本之间的两种关系：我们使用对比学习来利用样本对的信息以学习更好的目标表示，同时通过在邻域中聚合知识来细化伪标签。颜色表示伪标记的类别。0然而，TTA的挑战性设置引发了两个主要问题：1）如何在没有地面真值注释的情况下学习目标领域的表示；2）如何仅利用源域分类器作为源域的代理来构建目标域分类器。为了解决这些困难，现有的研究工作利用了图像/特征生成[23, 27]，类原型[28,57]，熵最小化[28,56]，自训练或伪标签[28]以及自监督辅助任务训练[50]。生成模型需要大量计算能力来生成目标风格的图像/特征[27]。基于熵最小化的方法具有竞争力，但直接优化熵会破坏目标上的模型校准。伪标签方法显示出有希望的结果，但其性能可能会受到噪声伪标签的影响[28]。测试时间训练[50]引入了一个自监督的辅助旋转预测任务，在源域和目标培训期间进行联合优化。这种方法的局限性在于它需要改变源培训协议，这对于所有感兴趣的模型可能不可行。此外，对比学习范式已被证明可以学习更具可转移性的表示，而不是作为预训练任务的旋转预测。最近，[55]在预训练阶段使用了自监督学习，然而，我们认为这种方法在适应阶段没有充分利用自监督表示学习的优势。在这项工