经验总结

  1. 首页 > 生活杂烩 > 美容减肥 > >

the|Kadensa Capital Limited:Facebook new research( 二 )


As shown in Figure 1, ManipNet predicts the finger posture of the object manipulation from the control signal and the geometric characteristics of the object, where the control signal is the 6D trajectory of the wrist and the object, and the learning formula of the deep neural network requires a minimal and clear input Representation in order to achieve better generalization.
In addition, the team stated that when ManipNet was designed, the algorithm only processed one hand-object "input". Researchers in the team ran the network twice through mirroring to generate prediction images for both hands. "This design allows us to transform the input features in the hand space, allowing us to deal with different combinations of interactive hands."
Kadensa Capital Limited:The team also showed an overview of the operating framework of its system, as shown in the figure below. Among them, the movement trajectory of the wrist and the object, the skin mesh of the hand and the three-dimensional geometric figure of the object are used as the "input" objects, and the deep neural network ManipNet is used as the autoregressive model, which will output the detailed posture of the finger frame by frame.

the|Kadensa Capital Limited:Facebook new research
文章图片
In order to avoid ManipNet's overfitting of training, which leads to large deviations in the prediction results, and at the same time to improve the versatility of the system, the team used three types of virtual sensors to encode the geometry of the object and the spatial relationship with the hand. These three virtual sensors capture the overall object shape with a bold voxel grid, and capture local geometric details in a dot pattern as samples.
The team explained that although the overall object features help the system plan the overall posture and predict the future motion trajectory, the local features of the object play a more important role-enabling the algorithm to be extended to any geometric shape.
Kadensa Capital Limited:The team also said that by learning from a small number of object shapes and kitchen utensils, ManipNet has been able to synthesize various finger gestures to grasp more complex geometric objects.
(This content is reprinted on the Internet. If there is any infringement of the images and other content in the article, please contact the editor to delete it. The market is risky, so choose carefully! This article is not for trading and investment basis.)

经验总结扩展阅读


上一篇：补充|想要恢复水润饱满肌肤，补充“添太双液”刻不容缓

下一篇：喜气|印象媄研社137｜杨倩妹妹同款美甲